Upload folder using huggingface_hub

.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

chat_template.jinja

@@ -0,0 +1,86 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# 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>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\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" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

config.json

@@ -0,0 +1,18 @@
+{
+  "auto_map": {
+    "AutoConfig": "configuration_qwen_vision.QwenVisionConfig",
+    "AutoModelForVision2Seq": "modeling_qwen_vision.QwenVisionForConditionalGeneration",
+    "AutoProcessor": "processing_qwen_vision.QwenVisionProcessor"
+  },
+  "clip_model_id": "openai/clip-vit-base-patch32",
+  "img_token": "[IMG]",
+  "img_token_count": 32,
+  "img_token_id": 151670,
+  "llm_hidden_size": 2560,
+  "llm_model_id": "Issactoto/qwen4b-instruct-cantone-ft",
+  "max_new_tokens": 256,
+  "model_type": "qwen_vision",
+  "projector_scale": 0.08,
+  "transformers_version": "5.0.0",
+  "vision_hidden_size": 768
+}

configuration_qwen_vision.py

@@ -0,0 +1,35 @@
+
+"""
+Configuration for QwenVision — a multimodal model combining
+openai/clip-vit-base-patch32 + linear projector + Qwen-based causal LLM.
+"""
+
+from transformers import PretrainedConfig
+
+
+class QwenVisionConfig(PretrainedConfig):
+    model_type = "qwen_vision"
+
+    def __init__(
+        self,
+        clip_model_id: str = "openai/clip-vit-base-patch32",
+        vision_hidden_size: int = 768,
+        img_token: str = "[IMG]",
+        img_token_id: int = -1,        # ← also add this (see Bug 2)
+        img_token_count: int = 32,
+        projector_scale: float = 0.08,
+        llm_model_id: str = "Issactoto/qwen4b-instruct-cantone-ft",
+        llm_hidden_size: int = 2560,
+        max_new_tokens: int = 256,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.clip_model_id = clip_model_id
+        self.vision_hidden_size = vision_hidden_size
+        self.img_token = img_token
+        self.img_token_id = img_token_id   # ← persist this
+        self.img_token_count = img_token_count
+        self.projector_scale = projector_scale
+        self.llm_model_id = llm_model_id
+        self.llm_hidden_size = llm_hidden_size
+        self.max_new_tokens = max_new_tokens

modeling_qwen_vision.py

@@ -0,0 +1,306 @@
+
+"""
+QwenVisionForConditionalGeneration
+-----------------------------------
+A HuggingFace-compatible Vision2Seq model combining:
+  - openai/clip-vit-base-patch32  (frozen visual encoder)
+  - Linear projector  (768 → LLM hidden size)
+  - Qwen-based causal LLM
+
+Weights layout on disk
+----------------------
+  config.json                  ← QwenVisionConfig
+  multimodal_adapter.pt        ← projector + img_embed  (your trained weights)
+  tokenizer files              ← saved alongside (includes [IMG] token)
+
+The LLM and CLIP are loaded from their HF hub IDs stored in config.json,
+so you don't need to re-upload those large weights.
+"""
+
+import os
+from typing import Optional, List, Union
+
+import torch
+import torch.nn as nn
+from transformers import (
+    PreTrainedModel,
+    AutoTokenizer,
+    AutoModelForCausalLM,
+    CLIPVisionModel,
+    CLIPImageProcessor,
+    BitsAndBytesConfig,
+    GenerationConfig,
+)
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+from configuration_qwen_vision import QwenVisionConfig
+
+
+class QwenVisionForConditionalGeneration(PreTrainedModel):
+    """
+    Multimodal model compatible with AutoModelForVision2Seq.
+
+    Usage
+    -----
+    model = QwenVisionForConditionalGeneration.from_pretrained(
+        "Issactoto/qwen4b-instruct-image-project"
+    )
+    """
+
+    config_class = QwenVisionConfig
+    # Tell HF which sub-modules carry their own configs (none here — we load by ID)
+    base_model_prefix = "llm"
+    supports_gradient_checkpointing = False
+
+    # ------------------------------------------------------------------ #
+    #  Construction                                                        #
+    # ------------------------------------------------------------------ #
+
+    def __init__(self, config: QwenVisionConfig, load_sub_models: bool = True):
+        super().__init__(config)
+        self.config = config
+
+        if load_sub_models:
+            self._build_vision_encoder()
+            self._build_llm()
+        # Projector and img_embed are always created (weights loaded later)
+        self._build_projector()
+
+    # ---- sub-model builders ------------------------------------------ #
+
+    def _build_vision_encoder(self):
+        self.vision = CLIPVisionModel.from_pretrained(
+            self.config.clip_model_id,
+            torch_dtype=torch.float16,
+        )
+        for p in self.vision.parameters():
+            p.requires_grad = False
+
+    def _build_llm(self, load_in_4bit: bool = True):
+        if load_in_4bit:
+            bnb_config = BitsAndBytesConfig(
+                load_in_4bit=True,
+                bnb_4bit_quant_type="nf4",
+                bnb_4bit_compute_dtype=torch.float16,
+            )
+        else:
+            bnb_config = None
+
+        self.llm = AutoModelForCausalLM.from_pretrained(
+            self.config.llm_model_id,
+            quantization_config=bnb_config,
+            device_map="auto",
+        )
+        for p in self.llm.parameters():
+            p.requires_grad = False
+
+        self.llm_dtype = self.llm.get_input_embeddings().weight.dtype
+
+    def _build_projector(self):
+        v_dim = self.config.vision_hidden_size  # 768
+        l_dim = self.config.llm_hidden_size     # 2560
+
+        # Infer dtype from LLM if available, else default float16
+        dtype = getattr(self, "llm_dtype", torch.float16)
+
+        self.projector = nn.Linear(v_dim, l_dim).to(dtype=dtype)
+        nn.init.normal_(self.projector.weight, std=0.02)
+
+        self.img_embed = nn.Parameter(torch.zeros(l_dim, dtype=dtype))
+
+    # ------------------------------------------------------------------ #
+    #  save_pretrained / from_pretrained hooks                            #
+    # ------------------------------------------------------------------ #
+
+    def save_pretrained(self, save_directory: str, **kwargs):
+        """
+        Saves:
+          - config.json
+          - multimodal_adapter.pt  (projector + img_embed only)
+
+        The LLM and CLIP are NOT re-saved; they are referenced by their
+        hub IDs in config.json and will be downloaded on load.
+        """
+        os.makedirs(save_directory, exist_ok=True)
+        self.config.save_pretrained(save_directory)
+
+        adapter_path = os.path.join(save_directory, "multimodal_adapter.pt")
+        torch.save(
+            {
+                "projector": self.projector.state_dict(),
+                "img_embed": self.img_embed.data,
+            },
+            adapter_path,
+        )
+        print(f"[QwenVision] Saved adapter weights → {adapter_path}")
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        pretrained_model_name_or_path: str,
+        load_in_4bit: bool = True,
+        **kwargs,
+    ):
+        """
+        Loads:
+          1. QwenVisionConfig from config.json
+          2. CLIP vision encoder  (from config.clip_model_id)
+          3. Qwen LLM             (from config.llm_model_id, 4-bit by default)
+          4. Projector + img_embed from multimodal_adapter.pt
+        """
+        # 1. Config
+        config = QwenVisionConfig.from_pretrained(
+            pretrained_model_name_or_path, **kwargs
+        )
+
+        # 2. Instantiate (this calls __init__ which builds sub-models)
+        model = cls(config, load_sub_models=True)
+
+        # Override 4-bit setting if caller passed it
+        if not load_in_4bit:
+            # Rebuild LLM without quantisation
+            model._build_llm(load_in_4bit=False)
+
+        # 3. Load adapter weights
+        adapter_path = os.path.join(
+            pretrained_model_name_or_path, "multimodal_adapter.pt"
+        )
+        if not os.path.isfile(adapter_path):
+            # Try downloading from hub
+            from huggingface_hub import hf_hub_download
+            adapter_path = hf_hub_download(
+                repo_id=pretrained_model_name_or_path,
+                filename="multimodal_adapter.pt",
+            )
+
+        device = next(model.llm.parameters()).device
+        ckpt = torch.load(adapter_path, map_location=device)
+        model.projector.load_state_dict(ckpt["projector"])
+        model.img_embed = nn.Parameter(ckpt["img_embed"].to(device))
+        print(f"[QwenVision] Loaded adapter weights from {adapter_path}")
+
+        return model
+
+    # ------------------------------------------------------------------ #
+    #  Forward                                                             #
+    # ------------------------------------------------------------------ #
+
+    def forward(
+        self,
+        pixel_values: torch.Tensor,
+        input_ids: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> CausalLMOutputWithPast:
+
+        device = input_ids.device
+        img_token_id = self.config.img_token_id  # set during processor init
+
+        # ---- 1. Encode image ----------------------------------------- #
+        with torch.no_grad():
+            vision_out = self.vision(pixel_values.to(next(self.vision.parameters()).device))
+            image_embed = vision_out.pooler_output.to(self.llm_dtype)
+
+        # ---- 2. Project → LLM space ---------------------------------- #
+        # [B, 1, L] → expand to [B, IMG_TOKEN_COUNT, L]
+        image_tokens = (
+            self.projector(image_embed)
+            .unsqueeze(1)
+            .expand(-1, self.config.img_token_count, -1)
+        )
+        image_tokens = image_tokens + self.img_embed.unsqueeze(0).unsqueeze(0)
+
+        # ---- 3. Build inputs_embeds with image injected -------------- #
+        inputs_embeds = self.llm.get_input_embeddings()(input_ids).clone()
+        mask = input_ids == img_token_id
+        b_idx, s_idx = mask.nonzero(as_tuple=True)
+        if b_idx.numel() > 0:
+            patch_idx = s_idx % self.config.img_token_count
+            inputs_embeds[b_idx, s_idx] = image_tokens[b_idx, patch_idx]
+
+        # ---- 4. LLM forward ------------------------------------------ #
+        return self.llm(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            labels=labels,
+        )
+
+    # ------------------------------------------------------------------ #
+    #  generate()  — wires into HF GenerationMixin                        #
+    # ------------------------------------------------------------------ #
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        pixel_values=None,
+        attention_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        **kwargs,
+    ):
+        """Called by model.generate() on each decoding step."""
+
+        # On the first step we inject image embeddings.
+        # On subsequent steps past_key_values is set, so we only pass the
+        # last token (standard autoregressive generation).
+
+        if past_key_values is not None:
+            # Subsequent decoding steps — just the new token
+            input_ids = input_ids[:, -1:]
+            return dict(
+                input_ids=input_ids,
+                pixel_values=None,        # already encoded
+                attention_mask=attention_mask,
+                past_key_values=past_key_values,
+                inputs_embeds=None,
+            )
+
+        # First step — build full inputs_embeds with image
+        if pixel_values is not None:
+            img_token_id = self.config.img_token_id
+            device = input_ids.device
+
+            with torch.no_grad():
+                vision_out = self.vision(pixel_values.to(next(self.vision.parameters()).device))
+                image_embed = vision_out.pooler_output.to(self.llm_dtype)
+
+            image_tokens = (
+                self.projector(image_embed)
+                .unsqueeze(1)
+                .expand(-1, self.config.img_token_count, -1)
+            )
+            image_tokens = image_tokens + self.img_embed.unsqueeze(0).unsqueeze(0)
+
+            inputs_embeds = self.llm.get_input_embeddings()(input_ids).clone()
+            mask = input_ids == img_token_id
+            b_idx, s_idx = mask.nonzero(as_tuple=True)
+            if b_idx.numel() > 0:
+                patch_idx = s_idx % self.config.img_token_count
+                inputs_embeds[b_idx, s_idx] = image_tokens[b_idx, patch_idx]
+
+            return dict(
+                input_ids=None,           # replaced by inputs_embeds
+                inputs_embeds=inputs_embeds,
+                pixel_values=None,
+                attention_mask=attention_mask,
+                past_key_values=past_key_values,
+            )
+
+        return dict(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+        )
+
+    def get_input_embeddings(self):
+        return self.llm.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.llm.set_input_embeddings(value)
+
+    def get_output_embeddings(self):
+        return self.llm.get_output_embeddings()
+
+    def can_generate(self):
+        return True

multimodal_adapter.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:03e4b9afd80d3e21e46f7b3b0d027e49a5dc6fe67b5ba14aa540f38fba23ccef
+size 3944708

processing_qwen_vision.py

@@ -0,0 +1,214 @@
+
+
+"""
+QwenVisionProcessor
+--------------------
+Wraps CLIPImageProcessor (for images) + Qwen tokenizer (for text) into a
+single AutoProcessor-compatible class.
+
+Supports apply_chat_template() so callers can use the exact same interface
+as granite-vision or LLaVA.
+"""
+
+import os
+from typing import List, Optional, Union
+from PIL import Image
+
+from transformers import (
+    ProcessorMixin,
+    CLIPImageProcessor,
+    AutoTokenizer,
+    BatchEncoding,
+)
+
+
+IMG_TOKEN = "[IMG]"
+IMG_TOKEN_COUNT = 32
+
+
+class QwenVisionProcessor(ProcessorMixin):
+    """
+    Processor for QwenVisionForConditionalGeneration.
+
+    Attributes exposed for AutoProcessor
+    -------------------------------------
+    attributes = ["image_processor", "tokenizer"]
+    """
+
+    # Required by ProcessorMixin / AutoProcessor registry
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "CLIPImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(self, image_processor: CLIPImageProcessor, tokenizer):
+        super().__init__(image_processor, tokenizer)
+        self.image_processor = image_processor
+        self.tokenizer = tokenizer
+
+        # Ensure [IMG] token exists
+        if tokenizer.convert_tokens_to_ids(IMG_TOKEN) == tokenizer.unk_token_id:
+            tokenizer.add_tokens([IMG_TOKEN])
+
+        self.img_token = IMG_TOKEN
+        self.img_token_id = tokenizer.convert_tokens_to_ids(IMG_TOKEN)
+        self.img_token_count = IMG_TOKEN_COUNT
+
+    # ------------------------------------------------------------------ #
+    #  Factory methods                                                     #
+    # ------------------------------------------------------------------ #
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: str, **kwargs):
+        image_processor = CLIPImageProcessor.from_pretrained(
+            "openai/clip-vit-base-patch32"
+        )
+        tokenizer = AutoTokenizer.from_pretrained(
+            pretrained_model_name_or_path, **kwargs
+        )
+        return cls(image_processor=image_processor, tokenizer=tokenizer)
+
+    def save_pretrained(self, save_directory: str, **kwargs):
+        os.makedirs(save_directory, exist_ok=True)
+        self.image_processor.save_pretrained(save_directory)
+        self.tokenizer.save_pretrained(save_directory)
+
+    # ------------------------------------------------------------------ #
+    #  apply_chat_template  — mirrors the granite-vision interface        #
+    # ------------------------------------------------------------------ #
+
+    def apply_chat_template(
+        self,
+        conversation: List[dict],
+        add_generation_prompt: bool = True,
+        tokenize: bool = True,
+        return_dict: bool = True,
+        return_tensors: Optional[str] = "pt",
+        images: Optional[List[Image.Image]] = None,
+        max_length: int = 512,
+        padding: Union[bool, str] = True,
+        truncation: bool = True,
+        enable_thinking: bool = False,
+        **kwargs,
+    ) -> BatchEncoding:
+        """
+        Parameters
+        ----------
+        conversation : list of dicts
+            Each dict has "role" and "content".
+            Content can be a string, or a list of dicts with "type" keys:
+              {"type": "image", "url": "/path/to/img.png"}
+              {"type": "text",  "text": "Your question"}
+        images : optional pre-loaded PIL images (overrides url extraction)
+        """
+
+        # ---- Extract images from conversation ------------------------ #
+        extracted_images: List[Image.Image] = []
+        text_messages = []
+
+        for turn in conversation:
+            role = turn["role"]
+            content = turn["content"]
+
+            if isinstance(content, str):
+                text_messages.append({"role": role, "content": content})
+                continue
+
+            # List of content blocks
+            text_parts = []
+            for block in content:
+                if block.get("type") == "image":
+                    # Load image from url/path if not supplied externally
+                    if images is None:
+                        url = block.get("url") or block.get("path")
+                        if url:
+                            extracted_images.append(Image.open(url).convert("RGB"))
+                    # Replace with [IMG] tokens placeholder
+                    img_placeholder = " ".join([self.img_token] * self.img_token_count)
+                    text_parts.append(img_placeholder)
+                elif block.get("type") == "text":
+                    text_parts.append(block["text"])
+
+            text_messages.append({"role": role, "content": " ".join(text_parts)})
+
+        # If caller provided images explicitly, use those
+        if images is not None:
+            extracted_images = images
+
+        # ---- Build prompt string via tokenizer's chat template ------- #
+        prompt_text = self.tokenizer.apply_chat_template(
+            text_messages,
+            tokenize=False,
+            add_generation_prompt=add_generation_prompt,
+            enable_thinking=enable_thinking,
+        )
+
+        if not tokenize:
+            return prompt_text  # type: ignore
+
+        # ---- Tokenise text ------------------------------------------- #
+        encoding = self.tokenizer(
+            prompt_text,
+            return_tensors=return_tensors,
+            padding=padding,
+            truncation=truncation,
+            max_length=max_length,
+            add_special_tokens=False,
+        )
+
+        # ---- Process images ------------------------------------------ #
+        if extracted_images:
+            pixel_values = self.image_processor(
+                images=extracted_images, return_tensors=return_tensors
+            )["pixel_values"]
+            encoding["pixel_values"] = pixel_values
+        else:
+            # No image supplied — caller must add pixel_values separately
+            pass
+
+        if return_dict:
+            return BatchEncoding(encoding)
+        return encoding
+
+    # ------------------------------------------------------------------ #
+    #  Standard __call__                                                   #
+    # ------------------------------------------------------------------ #
+
+    def __call__(
+        self,
+        text: Optional[Union[str, List[str]]] = None,
+        images: Optional[Union[Image.Image, List[Image.Image]]] = None,
+        return_tensors: Optional[str] = "pt",
+        padding: Union[bool, str] = True,
+        truncation: bool = True,
+        max_length: int = 512,
+        **kwargs,
+    ) -> BatchEncoding:
+
+        encoding = {}
+
+        if text is not None:
+            text_enc = self.tokenizer(
+                text,
+                return_tensors=return_tensors,
+                padding=padding,
+                truncation=truncation,
+                max_length=max_length,
+                **kwargs,
+            )
+            encoding.update(text_enc)
+
+        if images is not None:
+            if isinstance(images, Image.Image):
+                images = [images]
+            pixel_values = self.image_processor(
+                images=images, return_tensors=return_tensors
+            )["pixel_values"]
+            encoding["pixel_values"] = pixel_values
+
+        return BatchEncoding(encoding)
+
+    def decode(self, *args, **kwargs):
+        return self.tokenizer.decode(*args, **kwargs)
+
+    def batch_decode(self, *args, **kwargs):
+        return self.tokenizer.batch_decode(*args, **kwargs)

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a1519b02f3a5c1885951e21c729ecde6c737fd4b176697f41f4418bd9e7642c
+size 11423022

tokenizer_config.json

@@ -0,0 +1,15 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "is_local": false,
+  "model_max_length": 262144,
+  "pad_token": "<|PAD_TOKEN|>",
+  "padding_side": "left",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}