Upload 9 files

chat_template.jinja

@@ -0,0 +1,11 @@
+{% for message in messages %}
+{% if message['role'] == 'assistant' %}
+{{ '<|im_start|>assistant\n' }}
+{% generation %}
+{{ message['content'] }}
+{% endgeneration %}
+{{ '<|im_end|>' }}
+{% else %}
+{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>'}}
+{% endif %}
+{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}

config.json

@@ -0,0 +1,113 @@
+{
+  "architectures": [
+    "NeuroBLASTForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_neuroblast.NeuroBLASTConfig",
+    "AutoModelForCausalLM": "modeling_neuroblast.NeuroBLASTForCausalLM"
+  },
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "attention_every": 0,
+  "dropout": 0.0,
+  "dtype": "float32",
+  "hidden_act": "silu",
+  "hidden_size": 512,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "kernel_size": 5,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 32768,
+  "model_type": "neuroblast",
+  "num_associative_layers": 32,
+  "num_attention_heads": 16,
+  "num_hidden_layers": 72,
+  "num_key_value_heads": 8,
+  "num_motor_layers": 16,
+  "num_sensory_layers": 24,
+  "pad_token_id": 65537,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "scale": 1.0,
+  "sliding_window": null,
+  "temporal_kernel_size": 5,
+  "tie_word_embeddings": false,
+  "transformers_version": "4.57.1",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 65538
+}

configuration_neuroblast_jax.py

@@ -0,0 +1,322 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""NeuroBLASTConfig model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class NeuroBLASTConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen3Model`]. It is used to instantiate a
+    Qwen3 model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen3-8B [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 151936):
+            Vocabulary size of the Qwen3 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Qwen3Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 22016):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 32):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details, check out [this
+            paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `32`.
+        head_dim (`int`, *optional*, defaults to 128):
+            The attention head dimension.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`list[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`list[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers using full attention. The first `max_window_layers` layers will use full attention, while any
+            additional layer afterwards will use SWA (Sliding Window Attention).
+        layer_types (`list`, *optional*):
+            Attention pattern for each layer.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import Qwen3Model, Qwen3Config
+
+    >>> # Initializing a Qwen3 style configuration
+    >>> configuration = Qwen3Config()
+
+    >>> # Initializing a model from the Qwen3-8B style configuration
+    >>> model = Qwen3Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "frankqwenstein"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    # Default tensor parallel plan for base model `Qwen3`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        energy_dim=128,
+        intermediate_size=22016,
+        num_hidden_layers=32,
+        num_associative_layers=16,
+        num_sensory_layers=8,
+        num_motor_layers=8,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        head_dim=128,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        layer_types=None,
+        attention_dropout=0.0,
+        attention_every=0,
+        moe_topk=1,
+        moe_experts=1,
+        scale=1.0,
+        chunk_height=1,
+        chunk_top_k=2,
+        kernel_size=5,
+        temporal_kernel_size=5,  # New parameter for temporal convolution kernel size
+        dropout=0.0,
+        attn_dropout=0.0,
+        num_blocks=3,
+        conv_stride=1,  # New parameter for convolution stride
+        yarn_original_max_position_embeddings=8192,  # Original max position for YaRN
+        yarn_extrapolation_factor=1.0,  # YaRN extrapolation factor
+        yarn_attention_factor=1.0,  # YaRN attention factor
+        yarn_beta_fast=32,  # YaRN beta fast parameter
+        yarn_beta_slow=1,  # YaRN beta slow parameter
+        bidirectional_training=False,  # Enable bidirectional token prediction training
+        # Standard MoE parameters
+        use_moe=False,
+        num_experts=8,
+        num_experts_per_tok=2,
+        moe_intermediate_size=2048,
+        norm_topk_prob=True,
+        # Optimization parameters
+        capacity_factor=1.25,
+        eval_capacity_factor=2.0,
+        load_balancing_loss_coef=0.01,
+        expert_dropout=0.1,
+        routing_noise_std=0.1,
+        # Advanced features
+        num_gate_heads=1,
+        gating_strategy="standard",
+        expert_parallel=False,
+        track_expert_usage=True,
+        # Hierarchical MoE
+        num_expert_groups=4,
+        use_silu_after_blocks=False,
+        query_feature_map: str = "relu_squared",
+        kv_feature_map: str = "softplus",
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.energy_dim = energy_dim
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+
+        self.num_associative_layers = num_associative_layers
+        self.num_sensory_layers = num_sensory_layers
+
+        self.num_motor_layers = num_motor_layers
+
+        if (
+            num_hidden_layers
+            != num_associative_layers + num_sensory_layers + num_motor_layers
+        ):
+            self.num_hidden_layers = (
+                num_associative_layers + num_sensory_layers + num_motor_layers
+            )
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window if self.use_sliding_window else None
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.attention_every = attention_every
+        self.scale = scale
+        self.chunk_height = chunk_height
+        self.chunk_top_k = chunk_top_k
+        self.kernel_size = kernel_size
+        self.temporal_kernel_size = temporal_kernel_size
+        self.num_blocks = num_blocks
+        self.dropout = dropout
+        self.attn_dropout = attn_dropout
+        self.conv_stride = conv_stride
+        self.yarn_original_max_position_embeddings = (
+            yarn_original_max_position_embeddings
+        )
+        self.yarn_extrapolation_factor = yarn_extrapolation_factor
+        self.yarn_attention_factor = yarn_attention_factor
+        self.yarn_beta_fast = yarn_beta_fast
+        self.yarn_beta_slow = yarn_beta_slow
+        self.bidirectional_training = bidirectional_training
+        self.moe_topk = moe_topk
+        self.moe_experts = moe_experts
+
+        self.use_moe = use_moe
+        self.num_experts = num_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.moe_intermediate_size = moe_intermediate_size
+        self.norm_topk_prob = norm_topk_prob
+
+        # Optimization parameters
+        self.capacity_factor = capacity_factor
+        self.eval_capacity_factor = eval_capacity_factor
+        self.load_balancing_loss_coef = load_balancing_loss_coef
+        self.expert_dropout = expert_dropout
+        self.routing_noise_std = routing_noise_std
+
+        # Advanced features
+        self.num_gate_heads = num_gate_heads
+        self.gating_strategy = gating_strategy
+        self.expert_parallel = expert_parallel
+        self.track_expert_usage = track_expert_usage
+        self.use_silu_after_blocks = use_silu_after_blocks
+
+        # Hierarchical MoE
+        self.num_expert_groups = num_expert_groups
+
+        # Linear Attention Feature Maps
+        self.query_feature_map = query_feature_map
+        self.kv_feature_map = kv_feature_map
+
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+
+        self.layer_types = layer_types
+        if self.layer_types is None:
+            self.layer_types = [
+                ("full_attention") for i in range(self.num_hidden_layers)
+            ]
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d85fd3448ffeb9bce6bea7672865f6cec4e0e224400a1c38de97199c8359dd7
+size 2386809408

modeling_neuroblast_jax.py

@@ -0,0 +1,691 @@
+import math
+from typing import Optional, Tuple, Any, Callable
+
+import jax
+import jax.numpy as jnp
+from flax import linen as nn
+from flax.core.frozen_dict import FrozenDict
+from flax.linen.attention import dot_product_attention
+from flax.traverse_util import flatten_dict, unflatten_dict
+from transformers.modeling_flax_utils import FlaxPreTrainedModel
+from transformers.utils import logging
+
+from .configuration_neuroblast_jax import NeuroBLASTConfig
+
+logger = logging.get_logger(__name__)
+
+
+class NeuroBLASTRMSNorm(nn.Module):
+    hidden_size: int = 0
+    eps: float = 1e-6
+    dtype: Any = jnp.float32
+
+    def setup(self):
+        self.weight = self.param(
+            "weight",
+            lambda rng, shape: jnp.ones(shape, dtype=self.dtype),
+            (self.hidden_size,),
+        )
+
+    def __call__(self, hidden_states):
+        variance = jnp.mean(jnp.square(hidden_states), axis=-1, keepdims=True)
+        hidden_states = hidden_states * jax.lax.rsqrt(variance + self.eps)
+        return self.weight * hidden_states
+
+
+class NeuroBLASTMLP(nn.Module):
+    config: Optional[NeuroBLASTConfig] = None
+    dtype: Any = jnp.float32
+
+    def setup(self):
+        self.hidden_size = self.config.hidden_size
+        self.intermediate_size = self.config.intermediate_size
+        
+        self.gate_proj = nn.Dense(self.intermediate_size, use_bias=False, dtype=self.dtype)
+        self.up_proj = nn.Dense(self.intermediate_size, use_bias=False, dtype=self.dtype)
+        self.down_proj = nn.Dense(self.hidden_size, use_bias=False, dtype=self.dtype)
+        
+        # Activation function
+        if self.config.hidden_act == "silu":
+            self.act_fn = nn.silu
+        elif self.config.hidden_act == "gelu":
+            self.act_fn = nn.gelu
+        elif self.config.hidden_act == "relu":
+            self.act_fn = nn.relu
+        else:
+            raise ValueError(f"Unsupported activation: {self.config.hidden_act}")
+
+    def __call__(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+def rotate_half(x):
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return jnp.concatenate((-x2, x1), axis=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None):
+    
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class NeuroBLASTAttention(nn.Module):
+    config: Optional[NeuroBLASTConfig] = None
+    layer_idx: int = 0
+    use_rope: bool = True
+    dtype: Any = jnp.float32
+
+    def setup(self):
+        self.hidden_size = self.config.hidden_size
+        self.num_heads = self.config.num_attention_heads
+        self.head_dim = getattr(
+            self.config, "head_dim", self.hidden_size // self.num_heads
+        )
+        self.num_key_value_heads = self.config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attn_output_dim = self.num_heads * self.head_dim
+        
+        self.q_proj = nn.Dense(
+            self.attn_output_dim, 
+            use_bias=self.config.attention_bias, 
+            dtype=self.dtype
+        )
+        self.k_proj = nn.Dense(
+            self.num_key_value_heads * self.head_dim, 
+            use_bias=self.config.attention_bias, 
+            dtype=self.dtype
+        )
+        self.v_proj = nn.Dense(
+            self.num_key_value_heads * self.head_dim, 
+            use_bias=self.config.attention_bias, 
+            dtype=self.dtype
+        )
+        self.o_proj = nn.Dense(
+            self.hidden_size, 
+            use_bias=self.config.attention_bias, 
+            dtype=self.dtype
+        )
+        
+        self.q_norm = NeuroBLASTRMSNorm(self.head_dim, eps=self.config.rms_norm_eps, dtype=self.dtype)
+        self.k_norm = NeuroBLASTRMSNorm(self.head_dim, eps=self.config.rms_norm_eps, dtype=self.dtype)
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask=None,
+        position_embeddings=None,
+        deterministic: bool = True,
+    ):
+        batch_size, seq_len, _ = hidden_states.shape
+        
+        # Projection
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+        
+        # Reshape
+        query_states = query_states.reshape(batch_size, seq_len, self.num_heads, self.head_dim)
+        key_states = key_states.reshape(batch_size, seq_len, self.num_key_value_heads, self.head_dim)
+        value_states = value_states.reshape(batch_size, seq_len, self.num_key_value_heads, self.head_dim)
+        
+        # Norm
+        query_states = self.q_norm(query_states)
+        key_states = self.k_norm(key_states)
+        
+        # RoPE
+        if self.use_rope and position_embeddings is not None:
+            cos, sin = position_embeddings
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+            
+        if self.num_key_value_groups > 1:
+            key_states = jnp.repeat(key_states, self.num_key_value_groups, axis=2)
+            value_states = jnp.repeat(value_states, self.num_key_value_groups, axis=2)
+            
+        query_states = jnp.transpose(query_states, (0, 2, 1, 3))
+        key_states = jnp.transpose(key_states, (0, 2, 1, 3))
+        value_states = jnp.transpose(value_states, (0, 2, 1, 3))
+        
+        scale = 1.0 / jnp.sqrt(self.head_dim)
+        attn_weights = jnp.einsum("...qd,...kd->...qk", query_states, key_states) * scale
+        
+        if attention_mask is not None:
+            attn_weights = attn_weights + attention_mask
+            
+        attn_weights = nn.softmax(attn_weights, axis=-1)
+        
+        if self.config.attention_dropout > 0.0 and not deterministic:
+             attn_weights = nn.dropout(
+                 nn.make_rng(), 
+                 attn_weights, 
+                 deterministic=deterministic, 
+                 rate=self.config.attention_dropout
+             )
+             
+        attn_output = jnp.einsum("...qk,...kd->...qd", attn_weights, value_states)
+        
+        attn_output = jnp.transpose(attn_output, (0, 2, 1, 3))
+        attn_output = attn_output.reshape(batch_size, seq_len, self.attn_output_dim)
+        
+        attn_output = self.o_proj(attn_output)
+        
+        return attn_output
+
+
+
+class NeuroBLASTRMSNorm2d(nn.Module):
+    dim: int = 0
+    eps: float = 1e-6
+    dtype: Any = jnp.float32
+
+    def setup(self):
+        self.weight = self.param(
+            "weight",
+            lambda rng, shape: jnp.ones(shape, dtype=self.dtype),
+            (self.dim,),
+        )
+
+    def __call__(self, x):
+        variance = jnp.mean(jnp.square(x), axis=-1, keepdims=True)
+        x_norm = x * jax.lax.rsqrt(variance + self.eps)
+        return self.weight * x_norm
+
+
+class NeuroBLASTCausalConv2DBlock(nn.Module):
+    config: Optional[NeuroBLASTConfig] = None
+    dilation: int = 1
+    layer_idx: int = 0
+    dtype: Any = jnp.float32
+
+    def setup(self):
+        k = self.config.kernel_size
+        d = self.config.hidden_size
+        s = self.config.scale
+        
+        if s == 1:
+            self.conv = nn.Conv(
+                features=d,
+                kernel_size=(k, k),
+                kernel_dilation=(1, self.dilation),
+                padding=(k // 2, 0),
+                use_bias=False, 
+                dtype=self.dtype,
+            )
+            self.use_gating = False
+            self.use_projection = False
+        elif s > 1:
+            internal_dim = int(d * s)
+            self.conv = nn.Conv(
+                features=internal_dim,
+                kernel_size=(k, k),
+                kernel_dilation=(1, self.dilation),
+                padding=(k // 2, 0),
+                use_bias=False,
+                dtype=self.dtype,
+            )
+            self.use_gating = True
+            self.use_projection = False
+        else:
+            internal_dim = max(int(d * s), d // 4)
+            self.conv = nn.Conv(
+                features=internal_dim,
+                kernel_size=(k, k),
+                kernel_dilation=(1, self.dilation),
+                padding=(k // 2, 0),
+                use_bias=False,
+                dtype=self.dtype,
+            )
+            self.use_gating = False
+            self.use_projection = True
+            self.proj_back = nn.Conv(features=d, kernel_size=(1, 1), use_bias=False, dtype=self.dtype)
+
+        self.norm_in = NeuroBLASTRMSNorm2d(d, eps=self.config.rms_norm_eps, dtype=self.dtype)
+        self.norm_out = NeuroBLASTRMSNorm2d(d, eps=self.config.rms_norm_eps, dtype=self.dtype)
+        self.dropout = nn.Dropout(self.config.dropout)
+
+    def __call__(self, x, deterministic: bool = True):
+        B, H, W, C = x.shape  # Store original W for cropping!
+        residual = x
+        y = self.norm_in(x)
+        
+        k = self.config.kernel_size
+        pad_w = (k - 1) * self.dilation
+        
+        y_pad = jnp.pad(y, ((0, 0), (0, 0), (pad_w, 0), (0, 0)), mode='constant')
+        
+        y = self.conv(y_pad)
+        
+        y = y[:, :, -W:, :]
+        
+        if self.use_gating:
+            gate, val = jnp.split(y, 2, axis=-1)
+            y = val * nn.softmax(gate, axis=-1)
+        elif self.use_projection:
+            y = self.proj_back(y)
+            
+        y = self.norm_out(y)
+        
+        x = residual + self.dropout(y, deterministic=deterministic)
+        return x
+
+
+class NeuroBLASTDecoderLayer(nn.Module):
+    config: Optional[NeuroBLASTConfig] = None
+    layer_idx: int = 0
+    attention_type: str = "full_attention"
+    dtype: Any = jnp.float32
+
+    def setup(self):
+        self.hidden_size = self.config.hidden_size
+        
+        if self.attention_type == "linear_attention":
+            self.self_attn = NeuroBLASTLinearAttention(
+                config=self.config,
+                layer_idx=self.layer_idx,
+                query_feature_map_name=self.config.query_feature_map,
+                kv_feature_map_name=self.config.kv_feature_map,
+                dtype=self.dtype,
+            )
+        else:
+            self.self_attn = NeuroBLASTAttention(
+                config=self.config,
+                layer_idx=self.layer_idx,
+                use_rope=(self.attention_type != "no_rope"),
+                dtype=self.dtype,
+            )
+            
+        self.mlp = NeuroBLASTMLP(self.config, dtype=self.dtype)
+        self.input_layernorm = NeuroBLASTRMSNorm(self.hidden_size, eps=self.config.rms_norm_eps, dtype=self.dtype)
+        self.post_attention_layernorm = NeuroBLASTRMSNorm(self.hidden_size, eps=self.config.rms_norm_eps, dtype=self.dtype)
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask=None,
+        position_embeddings=None,
+        deterministic: bool = True,
+    ):
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        
+        hidden_states = self.self_attn(
+            hidden_states,
+            attention_mask=attention_mask,
+            position_embeddings=position_embeddings,
+            deterministic=deterministic,
+        )
+            
+        hidden_states = residual + hidden_states
+        
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        
+        return hidden_states
+
+
+class NeuroBLASTToken2D(nn.Module):
+    dtype: Any = jnp.float32
+
+    def __call__(self, x, mode="seq_to_2d"):
+        if mode == "seq_to_2d":
+            # x: (B, L, C) → (B, H, W, C) where H=1, W=L
+            # PyTorch: x.view(B, L, 1, C).permute(0, 3, 2, 1) → (B, C, 1, L) 
+            # In Flax channels-last: (B, H=1, W=L, C)
+            B, L, C = x.shape
+            x = x.reshape(B, 1, L, C)  # (B, H=1, W=L, C)
+            return x
+        else:
+            # x: (B, H, W, C) → (B, L, C) where L = W*H
+            # PyTorch: x.permute(0, 3, 2, 1).view(B, W * H, C)
+            #   Permute (B,C,H,W) → (B,W,H,C), then flatten to (B,W*H,C)
+            #   This means W varies first (causal ordering)
+            # 
+            # Flax: transpose (B, H, W, C) → (B, W, H, C), then flatten
+            B, H, W, C = x.shape
+            x = x.transpose(0, 2, 1, 3)  # (B, H, W, C) → (B, W, H, C) 
+            x = x.reshape(B, W * H, C)   # (B, W, H, C) → (B, W*H, C)
+            return x
+
+
+class NeuroBLASTRotaryEmbedding(nn.Module):
+    config: Optional[NeuroBLASTConfig] = None
+    dtype: Any = jnp.float32
+
+    def setup(self):
+        self.dim = self.config.head_dim
+        self.max_position_embeddings = self.config.max_position_embeddings
+        self.base = self.config.rope_theta
+        
+        # Precompute freqs
+        inv_freq = 1.0 / (self.base ** (jnp.arange(0, self.dim, 2, dtype=jnp.float32) / self.dim))
+        self.inv_freq = inv_freq
+
+    def __call__(self, x, position_ids):
+        # x: (B, L, H, D)
+        # position_ids: (B, L) or (1, L)
+        
+        inv_freq_expanded = self.inv_freq[None, :, None] # (1, D/2, 1)
+        
+        # position_ids: (B, L)
+        # We want (B, L, D/2)
+        
+        position_ids_expanded = position_ids[:, :, None] # (B, L, 1)
+        
+        # freqs: (B, L, D/2)
+        freqs = jnp.matmul(position_ids_expanded.astype(jnp.float32), self.inv_freq[None, None, :])
+        
+        # emb: (B, L, D)
+        emb = jnp.concatenate((freqs, freqs), axis=-1)
+        
+        cos = jnp.cos(emb)
+        sin = jnp.sin(emb)
+        
+        # Expand for heads: (B, L, 1, D)
+        cos = cos[:, :, None, :]
+        sin = sin[:, :, None, :]
+        
+        return cos, sin
+
+
+class NeuroBLASTModel(nn.Module):
+    config: Optional[NeuroBLASTConfig] = None
+    dtype: Any = jnp.float32
+
+    def setup(self):
+        self.embed_tokens = nn.Embed(
+            self.config.vocab_size,
+            self.config.hidden_size,
+            embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        
+        self.token2d = NeuroBLASTToken2D(dtype=self.dtype)
+        
+        sensory_layers = []
+        dilatation_step = 1
+        for i in range(self.config.num_sensory_layers):
+            if i % 2 == 0:
+                layer = NeuroBLASTDecoderLayer(
+                    self.config, layer_idx=i, attention_type="full_attention", dtype=self.dtype, name=f"sensory_layers_{i}"
+                )
+            else:
+                dilation = min(2 ** ((i - 1) // dilatation_step), 8)
+                layer = NeuroBLASTCausalConv2DBlock(
+                    self.config, dilation=dilation, layer_idx=i, dtype=self.dtype, name=f"sensory_layers_{i}"
+                )
+            sensory_layers.append(layer)
+        self.sensory_layers = sensory_layers
+            
+        self.sensory_to_associative = NeuroBLASTRMSNorm(self.config.hidden_size, eps=self.config.rms_norm_eps, dtype=self.dtype)
+        
+        associative_layers = []
+        next_layer_type = "full_attention"
+        for i in range(self.config.num_associative_layers):
+            idx = i + self.config.num_sensory_layers
+            if i % 2 == 0:
+                layer = NeuroBLASTDecoderLayer(
+                    self.config, layer_idx=idx, attention_type=next_layer_type, dtype=self.dtype, name=f"associative_layers_{i}"
+                )
+                if next_layer_type == "full_attention":
+                    next_layer_type = "no_rope"
+                else:
+                    next_layer_type = "full_attention"
+            else:
+                dilation = min(2 ** ((i - 1) // dilatation_step), 8)
+                layer = NeuroBLASTCausalConv2DBlock(
+                    self.config, dilation=dilation, layer_idx=idx, dtype=self.dtype, name=f"associative_layers_{i}"
+                )
+            associative_layers.append(layer)
+        self.associative_layers = associative_layers
+            
+        self.sensory_to_motor = NeuroBLASTRMSNorm(self.config.hidden_size, eps=self.config.rms_norm_eps, dtype=self.dtype)
+        
+        motor_layers = []
+        next_layer_type = "full_attention"
+        for i in range(self.config.num_motor_layers):
+            idx = i + self.config.num_sensory_layers + self.config.num_associative_layers
+            layer = NeuroBLASTDecoderLayer(
+                self.config, layer_idx=idx, attention_type=next_layer_type, dtype=self.dtype, name=f"motor_layers_{i}"
+            )
+            if next_layer_type == "full_attention":
+                next_layer_type = "no_rope"
+            else:
+                next_layer_type = "full_attention"
+            motor_layers.append(layer)
+        self.motor_layers = motor_layers
+            
+        self.norm = NeuroBLASTRMSNorm(self.config.hidden_size, eps=self.config.rms_norm_eps, dtype=self.dtype)
+        self.rotary_emb = NeuroBLASTRotaryEmbedding(self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        position_ids=None,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        batch_size, seq_len = input_ids.shape
+        
+        if position_ids is None:
+            position_ids = jnp.arange(seq_len, dtype="i4")[None, :]
+            
+        # Create attention mask
+        if attention_mask is None:
+            attention_mask = jnp.ones((batch_size, seq_len), dtype="i4")
+
+        # Build a boolean mask that enforces both padding and causality, then turn it
+        # into an additive bias for the attention logits.
+        attention_mask_bool = attention_mask.astype(bool)
+        causal_mask = nn.make_causal_mask(attention_mask_bool)
+        padding_mask = nn.make_attention_mask(attention_mask_bool, attention_mask_bool)
+        combined_mask = nn.combine_masks(causal_mask, padding_mask)
+        attention_bias = jnp.where(
+            combined_mask,
+            jnp.array(0.0, dtype=self.dtype),
+            jnp.array(jnp.finfo(self.dtype).min, dtype=self.dtype),
+        )
+        
+        # Embedding lookup
+        inputs_embeds = self.embed_tokens(input_ids)
+        
+        # Zero out padding token embeddings (equivalent to PyTorch's padding_idx)
+        # attention_mask is (B, L) with 1 for real tokens, 0 for padding
+        # Expand to (B, L, 1) to broadcast across hidden_size dimension
+        embedding_mask = attention_mask[:, :, None].astype(inputs_embeds.dtype)
+        inputs_embeds = inputs_embeds * embedding_mask
+        
+        hidden_states = inputs_embeds
+        
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+        
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+        
+        residual = hidden_states
+        
+        # Sensory
+        for i, layer in enumerate(self.sensory_layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+                
+            if i % 2 == 1:
+                hidden_states = self.token2d(hidden_states, mode="seq_to_2d")
+                hidden_states = layer(hidden_states, deterministic=deterministic)
+                hidden_states = self.token2d(hidden_states, mode="d2_to_seq")
+            else:
+                    hidden_states = layer(
+                        hidden_states,
+                        attention_mask=attention_bias,
+                        position_embeddings=position_embeddings,
+                        deterministic=deterministic,
+                    )
+                
+        hidden_states = hidden_states + self.sensory_to_associative(nn.silu(residual))
+        
+        # Associative
+        for i, layer in enumerate(self.associative_layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+                
+            if i % 2 == 1:
+                hidden_states = self.token2d(hidden_states, mode="seq_to_2d")
+                hidden_states = layer(hidden_states, deterministic=deterministic)
+                hidden_states = self.token2d(hidden_states, mode="d2_to_seq")
+            else:
+                    hidden_states = layer(
+                        hidden_states,
+                        attention_mask=attention_bias,
+                        position_embeddings=position_embeddings,
+                        deterministic=deterministic,
+                    )
+                
+        hidden_states = hidden_states + self.sensory_to_motor(nn.silu(-residual))
+        
+        # Motor
+        for i, layer in enumerate(self.motor_layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+                
+            hidden_states = layer(
+                hidden_states,
+                attention_mask=attention_bias,
+                position_embeddings=position_embeddings,
+                deterministic=deterministic,
+            )
+            
+        hidden_states = self.norm(hidden_states)
+        
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+            
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_attentions] if v is not None)
+            
+        return FlaxBaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+        )
+
+
+class NeuroBLASTForCausalLMModule(nn.Module):
+    config: Optional[NeuroBLASTConfig] = None
+    dtype: Any = jnp.float32
+
+    def setup(self):
+        self.model = NeuroBLASTModel(self.config, dtype=self.dtype)
+        self.lm_head = nn.Dense(self.config.vocab_size, use_bias=False, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        position_ids=None,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        
+        hidden_states = outputs[0] if not return_dict else outputs.last_hidden_state
+        logits = self.lm_head(hidden_states)
+        
+        if not return_dict:
+            return (logits,) + outputs[1:]
+            
+        return FlaxCausalLMOutput(
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+class NeuroBLASTForCausalLM(FlaxPreTrainedModel):
+    module_class = NeuroBLASTForCausalLMModule
+    config_class = NeuroBLASTConfig
+
+    def __init__(
+        self,
+        config: NeuroBLASTConfig,
+        input_shape: Tuple = (1, 1),
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        _do_init: bool = True,
+        **kwargs,
+    ):
+        module = NeuroBLASTForCausalLMModule(config=config, dtype=dtype, **kwargs)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
+        # init input tensors
+        input_ids = jnp.zeros(input_shape, dtype="i4")
+        attention_mask = jnp.ones_like(input_ids)
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_shape)
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        module_init_outputs = self.module.init(rngs, input_ids, attention_mask, position_ids, return_dict=False)
+
+        random_params = module_init_outputs["params"]
+
+        if params is not None:
+            random_params = flatten_dict(unflatten_dict(random_params) | unflatten_dict(params))
+            return FrozenDict(random_params)
+        return random_params
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        position_ids=None,
+        params: dict = None,
+        dropout_rng: jax.random.PRNGKey = None,
+        train: bool = False,
+        output_attentions: bool = None,
+        output_hidden_states: bool = None,
+        return_dict: bool = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if params is None:
+            params = self.params
+
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        return self.module.apply(
+            {"params": params or self.params},
+            input_ids=jnp.array(input_ids),
+            attention_mask=jnp.array(attention_mask) if attention_mask is not None else None,
+            position_ids=jnp.array(position_ids) if position_ids is not None else None,
+            deterministic=not train,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            rngs=rngs,
+        )
+
+from transformers.modeling_flax_outputs import FlaxBaseModelOutput, FlaxCausalLMOutput
+
+

special_tokens_map.json

@@ -0,0 +1,77 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<think>",
+    "</think>",
+    "source_1",
+    "source_2",
+    "source_3",
+    "source_4",
+    "source_5",
+    "source_6",
+    "source_7",
+    "source_8",
+    "source_9",
+    "source_10",
+    "<ref",
+    "</ref>",
+    "→",
+    "↺",
+    "※",
+    "?maybe?",
+    "●",
+    "◐",
+    "○",
+    "⚠",
+    "☐",
+    "☑",
+    "✓",
+    "⟨H≈0.1⟩",
+    "⟨H≈0.2⟩",
+    "⟨H≈0.3⟩",
+    "⟨H≈0.4⟩",
+    "⟨H≈0.5⟩",
+    "⟨H≈0.6⟩",
+    "⟨H≈0.7⟩",
+    "⟨H≈0.8⟩",
+    "⟨H≈0.9⟩",
+    "⟨H≈1.0⟩",
+    "⟨H≈1.1⟩",
+    "⟨H≈1.2⟩",
+    "⟨H≈1.3⟩",
+    "⟨H≈1.4⟩",
+    "⟨H≈1.5⟩",
+    "⟨H≈1.6⟩",
+    "⟨H≈1.7⟩",
+    "⟨H≈1.8⟩"
+  ],
+  "bos_token": {
+    "content": "<|im_start|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "<|mask|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|pad|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer_config.json

@@ -0,0 +1,474 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<|begin_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "<|end_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65491": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65492": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65493": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65494": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65495": {
+      "content": "source_1",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65496": {
+      "content": "source_2",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65497": {
+      "content": "source_3",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65498": {
+      "content": "source_4",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65499": {
+      "content": "source_5",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65500": {
+      "content": "source_6",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65501": {
+      "content": "source_7",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65502": {
+      "content": "source_8",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65503": {
+      "content": "source_9",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65504": {
+      "content": "source_10",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65505": {
+      "content": "<ref",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65506": {
+      "content": "</ref>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65507": {
+      "content": "→",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65508": {
+      "content": "↺",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65509": {
+      "content": "※",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65510": {
+      "content": "?maybe?",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65511": {
+      "content": "●",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65512": {
+      "content": "◐",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65513": {
+      "content": "○",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65514": {
+      "content": "⚠",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65515": {
+      "content": "☐",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65516": {
+      "content": "☑",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65517": {
+      "content": "✓",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65518": {
+      "content": "⟨H≈0.1⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65519": {
+      "content": "⟨H≈0.2⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65520": {
+      "content": "⟨H≈0.3⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65521": {
+      "content": "⟨H≈0.4⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65522": {
+      "content": "⟨H≈0.5⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65523": {
+      "content": "⟨H≈0.6⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65524": {
+      "content": "⟨H≈0.7⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65525": {
+      "content": "⟨H≈0.8⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65526": {
+      "content": "⟨H≈0.9⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65527": {
+      "content": "⟨H≈1.0⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65528": {
+      "content": "⟨H≈1.1⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65529": {
+      "content": "⟨H≈1.2⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65530": {
+      "content": "⟨H≈1.3⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65531": {
+      "content": "⟨H≈1.4⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65532": {
+      "content": "⟨H≈1.5⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65533": {
+      "content": "⟨H≈1.6⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65534": {
+      "content": "⟨H≈1.7⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65535": {
+      "content": "⟨H≈1.8⟩",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65536": {
+      "content": "<|mask|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "65537": {
+      "content": "<|pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<think>",
+    "</think>",
+    "source_1",
+    "source_2",
+    "source_3",
+    "source_4",
+    "source_5",
+    "source_6",
+    "source_7",
+    "source_8",
+    "source_9",
+    "source_10",
+    "<ref",
+    "</ref>",
+    "→",
+    "↺",
+    "※",
+    "?maybe?",
+    "●",
+    "◐",
+    "○",
+    "⚠",
+    "☐",
+    "☑",
+    "✓",
+    "⟨H≈0.1⟩",
+    "⟨H≈0.2⟩",
+    "⟨H≈0.3⟩",
+    "⟨H≈0.4⟩",
+    "⟨H≈0.5⟩",
+    "⟨H≈0.6⟩",
+    "⟨H≈0.7⟩",
+    "⟨H≈0.8⟩",
+    "⟨H≈0.9⟩",
+    "⟨H≈1.0⟩",
+    "⟨H≈1.1⟩",
+    "⟨H≈1.2⟩",
+    "⟨H≈1.3⟩",
+    "⟨H≈1.4⟩",
+    "⟨H≈1.5⟩",
+    "⟨H≈1.6⟩",
+    "⟨H≈1.7⟩",
+    "⟨H≈1.8⟩"
+  ],
+  "bos_token": "<|im_start|>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|im_end|>",
+  "extra_special_tokens": {},
+  "mask_token": "<|mask|>",
+  "max_length": 256,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_to_multiple_of": null,
+  "pad_token": "<|pad|>",
+  "pad_token_type_id": 0,
+  "padding_side": "right",
+  "stride": 0,
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "truncation_side": "right",
+  "truncation_strategy": "longest_first"
+}