Upload 14 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,150 @@
 ---
 license: mit
+tags:
+
+  - sentence-transformers
+  - sentence-similarity
+  - mteb
+  - retriever
+  - text-embeddings-inference
+
 ---
+
+# Jasper-Token-Compression-V2
+
+## Introduction
+
+Inspired by Deepseek-OCR, this is the first vector model in the Jasper and Stella series to use **dynamic text token
+compression**. Through the combination of vector distillation and contrastive learning, our model can compress text by
+3x
+while still achieving excellent performance!
+
+## Features
+
+- ⭐⭐⭐ Supports bilingual (Chinese and English)
+- ⭐⭐⭐⭐⭐⭐ Dynamic token compression - tested to achieve excellent results even when compressing text to 0.33x of original
+  length
+- ⭐⭐⭐ Combines vector distillation with contrastive learning to further improve performance on retrieval tasks
+- ⭐⭐ 12 million unsupervised data distillation
+- ⭐⭐ 0.6B parameter size
+
+## Technical Details
+
+### Dynamic Text Token Compression
+
+My implementation is very simple: After text passes through the `word_embedding` layer, it immediately goes through a
+`Qwen3MLP` (approximately 3 fully connected layers), then I calculate the compressed length, and finally use
+`adaptive_avg_pool1d` to compress tokens to that length.
+
+The compression length calculation logic is as follows:
+
+```python
+real_length = 1000  # Actual token count of the text
+length_threshold = 80  # Compress only if exceeding this threshold
+compression_ratio = 0.333
+if real_length <= length_threshold:
+    # No compression
+    pass
+else:
+    target_length = int(length_threshold + (real_length - length_threshold) * compression_ratio)
+```
+
+For implementation details, please refer to the `modeling_qwen3_jasper.py` file in this directory.
+
+### Vector Distillation + Contrastive Learning
+
+First, we compute teacher vectors for texts in the contrastive learning training set, then use the following three
+losses during training:
+
+1. Cosine loss: Standard vector distillation loss
+2. InfoNCE (hard loss): Standard contrastive learning loss function
+3. KL divergence (soft loss): KL divergence between student score matrix and teacher score matrix. The score matrix is
+   the scores between query and all documents(i.e. positive doc, hard negative docs, other in-batch docs).
+
+#### Evaluation
+
+My prompt strategy and specific content are consistent with the QZhou model. Please refer to their evaluation
+script: https://github.com/Kingsoft-LLM/QZhou-Embedding
+
+### Usage
+
+```py
+import torch
+from sentence_transformers import SentenceTransformer
+
+if __name__ == "__main__":
+    model_name_or_path = "infgrad/Jasper-Token-Compression-V2"
+    model = SentenceTransformer(
+        model_name_or_path,
+        model_kwargs={
+            "torch_dtype": torch.bfloat16,
+            "attn_implementation": "sdpa",  # We support flash_attention_2; sdpa; eager
+            "trust_remote_code": True
+        },
+        trust_remote_code=True,
+        tokenizer_kwargs={"padding_side": "left"},
+        device="cpu",
+    )
+
+    queries = [
+        "What is photosynthesis?",
+        "Who invented the telephone?",
+    ]
+    documents = [
+        "Photosynthesis is the process by which green plants use sunlight, carbon dioxide, and water to produce glucose and oxygen",
+        "Alexander Graham Bell is credited with inventing the first practical telephone in 1876, receiving US patent number 174,465 for his device."
+    ]
+    # The smaller the compression_ratio parameter, the faster the speed, but the quality will correspondingly decrease.
+    # Based on our parameter settings during training and test results, we recommend a range between 0.3-0.8.
+    query_embeddings = model.encode(queries, prompt_name="query", normalize_embeddings=True, compression_ratio=0.3333)
+    document_embeddings = model.encode(documents, normalize_embeddings=True, compression_ratio=0.3333)
+
+    similarity = model.similarity(query_embeddings, document_embeddings)
+    print(similarity)
+
+```
+
+### Limitations and TODO
+
+#### Retrieval performance
+
+I found that distilled models struggle to approach the retrieval performance of teacher models, which is why I
+specifically used contrastive learning + distillation learning to enhance the student model. However, I found that while
+the enhanced model showed improvement on retrieval test sets, there is still a significant gap compared to mainstream
+models.
+**Therefore, I believe that how to improve the retrieval performance of distilled models is a very necessary and
+valuable
+research direction.**
+
+#### More reasonable text token compression modules
+
+There is limited research on text token compression currently, and I have only tried the simplest approach. I believe
+more reasonable text compression modules can definitely be found.
+
+#### Text length
+
+I only distilled texts up to 1024 tokens in length, so there should be performance degradation when text length exceeds
+1024.
+
+### Citation
+
+If you find our work worth citing, please use the following citation.
+
+For distillation, please cite the following paper:
+**Jasper and Stella Technical Report:**
+
+```
+
+@misc{zhang2025jasperstelladistillationsota,
+      title={Jasper and Stella: distillation of SOTA embedding models}, 
+      author={Dun Zhang and Jiacheng Li and Ziyang Zeng and Fulong Wang},
+      year={2025},
+      eprint={2412.19048},
+      archivePrefix={arXiv},
+      primaryClass={cs.IR},
+      url={https://arxiv.org/abs/2412.19048}, 
+}
+
+```
+
+For text compression, please cite this link directly. We will consider writing a report later.

config.json

@@ -0,0 +1,33 @@
+{
+  "architectures": [
+    "JasperV2Encoder"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoModel": "modeling_qwen3_jasper.JasperV2Encoder"
+  },
+  "bos_token_id": 151643,
+  "dtype": "bfloat16",
+  "eos_token_id": 151643,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "transformers_version": "4.57.1",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "vocab_size": 151669
+}

config_sentence_transformers.json

@@ -0,0 +1,8 @@
+{
+  "prompts": {
+    "query": "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: ",
+    "document": ""
+  },
+  "default_prompt_name": null,
+  "similarity_fn_name": "cosine"
+}

configuration.json

@@ -0,0 +1,5 @@
+{
+  "framework": "pytorch",
+  "task": "text-generation",
+  "allow_remote": true
+}

custom_st.py

@@ -0,0 +1,9 @@
+import torch
+from sentence_transformers.models import Transformer as BaseTransformer
+
+
+class JasperTransformer(BaseTransformer):
+    def forward(self, features: dict[str, torch.Tensor], **kwargs) -> dict[str, torch.Tensor]:
+        vectors = self.auto_model(**features, **kwargs)
+        features.update({"sentence_embedding": vectors})
+        return features

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "max_new_tokens": 2048,
+  "transformers_version": "4.51.3"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4a7b9e62fcbfcfbffe217d4cf45b4957502df1bce4fbc05a8f13c1f05f33158
+size 1214661536

modeling_qwen3_jasper.py

@@ -0,0 +1,146 @@
+import random
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import Qwen3PreTrainedModel, Qwen3Config, Qwen3Model
+from transformers.models.qwen3.modeling_qwen3 import Qwen3MLP
+
+
+class TokenCompressor(nn.Module):
+    """
+    自适应Token压缩模块
+    对于长度超过阈值的序列，使用adaptive_avg_pool1d进行压缩
+    压缩后长度 = 阈值 + 超出部分 * compression_ratio
+    """
+
+    def __init__(self, length_threshold: int = 512, compression_ratio: float = 0.3):
+        super().__init__()
+        self.length_threshold = length_threshold
+        self.compression_ratio = compression_ratio
+
+    def forward(
+            self, token_embeddings: torch.Tensor, attention_mask: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """
+        对token embeddings进行自适应压缩
+        Args:
+            token_embeddings: [batch_size, seq_len, hidden_size]
+            attention_mask: [batch_size, seq_len]
+        Returns:
+            compressed_embeddings: 压缩后的embeddings
+            compressed_mask: 压缩后的attention mask
+        """
+        padding_side = 'right' if (attention_mask[:, -1] == 0).any() else 'left'
+
+        compressed_embeddings_list = []
+        compressed_masks_list = []
+        for text_idx in range(token_embeddings.shape[0]):
+            # 获取当前样本的有效长度
+            real_length = int(attention_mask[text_idx].sum().item())
+            if real_length <= self.length_threshold:
+                # 根据padding方向提取有效的token embeddings
+                if padding_side == 'left':
+                    # 左填充：有效tokens在右边
+                    valid_embeddings = token_embeddings[text_idx:text_idx + 1, -real_length:, :]
+                else:
+                    # 右填充：有效tokens在左边
+                    valid_embeddings = token_embeddings[text_idx:text_idx + 1, :real_length, :]
+                compressed_embeddings_list.append(valid_embeddings)
+                compressed_masks_list.append([1] * real_length)
+            else:
+                target_length = int(
+                    self.length_threshold + (real_length - self.length_threshold) * self.compression_ratio
+                )
+                # 根据padding方向提取有效的token embeddings
+                if padding_side == 'left':
+                    # 左填充：有效tokens在右边
+                    valid_embeddings = token_embeddings[text_idx:text_idx + 1, -real_length:, :]
+                else:
+                    # 右填充：有效tokens在左边
+                    valid_embeddings = token_embeddings[text_idx:text_idx + 1, :real_length, :]
+
+                # 使用adaptive_avg_pool1d进行压缩
+                compressed_embeddings_list.append(
+                    F.adaptive_avg_pool1d(
+                        valid_embeddings.transpose(1, 2), target_length
+                    ).transpose(1, 2)
+                )
+                # print("valid_embeddings.shape,target_length,compressed_embeddings_list[-1].shape",valid_embeddings.shape,target_length,compressed_embeddings_list[-1].shape)
+                compressed_masks_list.append([1] * target_length)
+
+        # 重新组合为token_embeddings和attention_mask
+        new_seq_len = max((len(_mask) for _mask in compressed_masks_list))
+        new_attention_mask = torch.tensor(
+            [
+                _mask + [0] * (new_seq_len - len(_mask))
+                if padding_side == "right"
+                else
+                [0] * (new_seq_len - len(_mask)) + _mask
+                for _mask in compressed_masks_list
+            ],
+            dtype=torch.long,
+            device=token_embeddings.device
+        )
+
+        # 生成新的token_embeddings
+        batch_size = token_embeddings.shape[0]
+        hidden_size = token_embeddings.shape[2]
+        new_token_embeddings = torch.zeros(
+            batch_size, new_seq_len, hidden_size,
+            dtype=token_embeddings.dtype,
+            device=token_embeddings.device
+        )
+
+        for idx, compressed_emb in enumerate(compressed_embeddings_list):
+            seq_len = compressed_emb.shape[1]
+            if padding_side == "right":
+                new_token_embeddings[idx, :seq_len, :] = compressed_emb.squeeze(0)
+            else:
+                # print("new_token_embeddings.shape,compressed_emb.shape",new_token_embeddings.shape,compressed_emb.shape)
+                new_token_embeddings[idx, -seq_len:, :] = compressed_emb.squeeze(0)
+
+        return new_token_embeddings, new_attention_mask
+
+
+
+class JasperV2Encoder(Qwen3PreTrainedModel):
+
+    def __init__(self, config: Qwen3Config):
+        super().__init__(config)
+        self.model = Qwen3Model(config)
+        self.jasper_mlp = Qwen3MLP(config=config)
+        self.linear_1 = nn.Linear(in_features=config.hidden_size, out_features=2048, bias=True)
+        self.token_compressor = TokenCompressor(length_threshold=80, compression_ratio=0.5)
+        self.post_init()
+
+    def forward(
+            self,
+            input_ids: torch.Tensor,
+            attention_mask: torch.Tensor,
+            *args,
+            **kwargs
+    ) -> torch.Tensor:
+        # token_embeddings.shape batch_size*seq_len*hidden_size
+        token_embeddings = self.model.embed_tokens(input_ids)
+        token_embeddings = self.jasper_mlp(token_embeddings)
+
+        self.token_compressor.compression_ratio = kwargs.get(
+            "compression_ratio",
+            self.token_compressor.compression_ratio
+        )
+        compressed_token_embeddings, attention_mask = self.token_compressor(token_embeddings, attention_mask)
+        compressed_token_embeddings = self.model(
+            inputs_embeds=compressed_token_embeddings, attention_mask=attention_mask
+        )["last_hidden_state"]
+
+        # 生成句向量
+        input_mask_expanded = (
+            attention_mask.unsqueeze(-1).expand(compressed_token_embeddings.size()).to(
+                compressed_token_embeddings.dtype)
+        )
+        sum_embeddings = torch.sum(compressed_token_embeddings * input_mask_expanded, 1)
+        sum_mask = input_mask_expanded.sum(1)
+        sum_mask = torch.clamp(sum_mask, min=1e-9)
+        vector = sum_embeddings / sum_mask
+        return self.linear_1(vector)

modules.json

@@ -0,0 +1,17 @@
+[
+  {
+    "idx": 0,
+    "name": "0",
+    "path": "",
+    "type": "custom_st.JasperTransformer",
+    "kwargs": [
+      "compression_ratio"
+    ]
+  },
+  {
+    "idx": 1,
+    "name": "1",
+    "path": "1_Normalize",
+    "type": "sentence_transformers.models.Normalize"
+  }
+]

sentence_bert_config.json

@@ -0,0 +1,7 @@
+{
+    "max_seq_length": 32768,
+    "do_lower_case": false,
+    "tokenizer_args": {
+        "padding_side": "left"
+    }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:def76fb086971c7867b829c23a26261e38d9d74e02139253b38aeb9df8b4b50a
+size 11423705

tokenizer_config.json

@@ -0,0 +1,240 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0].role == 'system' %}\n        {{- messages[0].content + '\\n\\n' }}\n    {%- endif %}\n    {{- \"# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0].role == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0].content + '<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n    {%- set index = (messages|length - 1) - loop.index0 %}\n    {%- if ns.multi_step_tool and message.role == \"user\" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}\n        {%- set ns.multi_step_tool = false %}\n        {%- set ns.last_query_index = index %}\n    {%- endif %}\n{%- endfor %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {%- set content = message.content %}\n        {%- set reasoning_content = '' %}\n        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}\n            {%- set reasoning_content = message.reasoning_content %}\n        {%- else %}\n            {%- if '</think>' in message.content %}\n                {%- set content = message.content.split('</think>')[-1].lstrip('\\n') %}\n                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n            {%- endif %}\n        {%- endif %}\n        {%- if loop.index0 > ns.last_query_index %}\n            {%- if loop.last or (not loop.last and reasoning_content) %}\n                {{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content.strip('\\n') + '\\n</think>\\n\\n' + content.lstrip('\\n') }}\n            {%- else %}\n                {{- '<|im_start|>' + message.role + '\\n' + content }}\n            {%- endif %}\n        {%- else %}\n            {{- '<|im_start|>' + message.role + '\\n' + content }}\n        {%- endif %}\n        {%- if message.tool_calls %}\n            {%- for tool_call in message.tool_calls %}\n                {%- if (loop.first and content) or (not loop.first) %}\n                    {{- '\\n' }}\n                {%- endif %}\n                {%- if tool_call.function %}\n                    {%- set tool_call = tool_call.function %}\n                {%- endif %}\n                {{- '<tool_call>\\n{\"name\": \"' }}\n                {{- tool_call.name }}\n                {{- '\", \"arguments\": ' }}\n                {%- if tool_call.arguments is string %}\n                    {{- tool_call.arguments }}\n                {%- else %}\n                    {{- tool_call.arguments | tojson }}\n                {%- endif %}\n                {{- '}\\n</tool_call>' }}\n            {%- endfor %}\n        {%- endif %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if loop.first or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n    {%- if enable_thinking is defined and enable_thinking is false %}\n        {{- '<think>\\n\\n</think>\\n\\n' }}\n    {%- endif %}\n{%- endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}