update model weights

.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 +1,163 @@
 ---
 license: apache-2.0
+datasets:
+- TIGER-Lab/MMEB-train
+language:
+- en
+metrics:
+- precision
+base_model:
+- Qwen/Qwen2-VL-7B-Instruct
+pipeline_tag: sentence-similarity
+library_name: transformers
+tags:
+- Qwen2-VL
+- qwen2-vl
+- MMEB
 ---
+
+<p align="center">
+  <img src="images/SaHa_logo.png" alt="SaHa Logo" style="width: 100%; max-width: 450px;">
+</p>
+
+# SaHa-Qwen2-VL-7B-Instruct
+
+## Model Summary
+
+**SaHa-Qwen2-VL-7B-Instruct** is a state-of-the-art universal multimodal embedding model based on the **Qwen2-VL-7B-Instruct** architecture. This model has been fine-tuned using our innovative Self-aware Hard Negative Sampling (SaHa) strategy, which is designed to efficiently adapt generative Multimodal Large Language Models (MLLMs) for discriminative embedding tasks.
+
+Our approach leverages a hierarchical embedding prompt to unlock the powerful zero-shot capabilities of MLLMs and then fine-tunes the model with SaHa to achieve superior performance on universal multimodal retrieval benchmarks. This model significantly reduces the computational costs associated with traditional contrastive pre-training while delivering state-of-the-art results.
+
+For more details, please refer to our paper: [From Generator to Embedder: Harnessing Innate Abilities of Multimodal LLMs via Building Zero-Shot Discriminative Embedding Model](https://arxiv.org/abs/2508.00955) and our [GitHub repository](https://github.com/yeongjoonJu/Gen2Embed).
+
+## How to Use
+
+You can easily use this model with the `transformers` library for sentence and image similarity tasks. Make sure you have the latest version of `transformers`, `torch`, and `Pillow` installed.
+
+```bash
+pip install transformers>=4.46.1 torch pillow
+```
+
+### Get Embeddings from Text or Image
+
+Here's how to get embeddings for text or image inputs. The model uses a specific prompt structure to generate high-quality embeddings.
+
+**Load Model**
+
+```python
+import torch
+from transformers import AutoProcessor, AutoConfig, Qwen2VLForConditionalGeneration
+
+# Load the model and tokenizer
+model_id = "Y-J-Ju/SaHa-Qwen2-VL-7B-Instruct"
+
+config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
+config._attn_implementation = "flash_attention_2"
+config.vision_config._attn_implementation = "flash_attention_2"
+
+model = Qwen2VLForConditionalGeneration.from_pretrained(
+    model_id, torch_dtype=torch.bfloat16, config=config, device_map="cuda:0"
+)
+processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True,
+                                          min_pixels=256 * 28 * 28, max_pixels=1280 * 28 * 28)
+```
+
+**Data Preparation and Prompting**
+```python
+texts = [
+    "The Tesla Cybertruck is a battery electric pickup truck built by Tesla, Inc. since 2023.",
+    "Korea University",
+]
+images = [
+    'https://upload.wikimedia.org/wikipedia/commons/e/e9/Tesla_Cybertruck_damaged_window.jpg',
+    'https://upload.wikimedia.org/wikipedia/commons/thumb/7/74/Korea_University.jpg/960px-Korea_University.jpg',
+]
+task_instruction = 'Find an image that matches the given text.'
+
+system_prompt = "Given an image, summarize the provided image in one word. Given only text, describe the text in one word."
+represent_prompt = "Represent the given text in one word."
+
+query_form = '<|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\n{task_instruction}\n{query}\n{represent_prompt}<|im_end|>\n<|im_start|>assistant\n'
+candidate_form = '<|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\n{cand}<|im_end|>\n<|im_start|>assistant\n'
+
+queries = [
+    query_form.format(system_prompt=system_prompt, task_instruction=task_instruction, query=text, represent_prompt=represent_prompt)
+    for text in texts
+]
+candidates = [
+    candidate_form.format(system_prompt=system_prompt, cand='<|image_pad|>')
+    for _ in images
+]
+```
+
+**Get Embeddings**
+```python
+from PIL import Image
+import io
+from urllib import request
+import torch.nn.functional as F
+
+## Query (Text)
+inputs = processor(text=queries, images=None, return_tensors="pt", padding=True)
+model_input = {k: v if isinstance(v, list) else v.to(model.device) for k, v in inputs.items()}
+outputs = model(**model_input, return_dict=True, output_hidden_states=True)
+hidden_states = outputs.hidden_states[-1]
+query_embed = hidden_states[:,-1]
+
+## Candidate (Image)
+pil_images = [Image.open(io.BytesIO(request.urlopen(url).read())) for url in images]
+inputs = processor(text=candidates, images=pil_images, return_tensors="pt", padding=True)
+model_input = {k: v if isinstance(v, list) else v.to(model.device) for k, v in inputs.items()}
+outputs = model(**model_input, return_dict=True, output_hidden_states=True)
+cand_embed = outputs.hidden_states[-1][:,-1]
+
+query_embed = F.normalize(query_embed, p=2, dim=-1)
+cand_embed = F.normalize(cand_embed, p=2, dim=-1)
+print(query_embed @ cand_embed.T)
+```
+
+**Outputs (Similarity)**
+
+~~~python
+tensor([[ 0.3848, -0.0197],
+        [-0.0221,  0.2949]], device='cuda:0', dtype=torch.bfloat16)
+~~~
+
+## Training and Evaluation
+
+### Training Data
+
+The model was fine-tuned on the **Massive Multimodal Embedding Benchmark (MMEB)** training set, which consists of approximately 829,000 pairs from 20 in-domain datasets.
+
+  * **Training Data:** [TIGER-Lab/MMEB-train](https://huggingface.co/datasets/TIGER-Lab/MMEB-train)
+
+### Evaluation Data
+
+The model's performance was evaluated on the MMEB evaluation set, which includes 36 datasets covering four meta-tasks: Classification, Visual Question Answering (VQA), Retrieval, and Visual Grounding.
+
+  * **Evaluation Data:** [TIGER-Lab/MMEB-eval](https://huggingface.co/datasets/TIGER-Lab/MMEB-eval)
+
+### Performance
+
+The SaHa-Qwen2-VL-7B-Instruct model achieves state-of-the-art performance in its parameter class on the MMEB benchmark, outperforming methods that rely on large-scale contrastive pre-training.
+
+| Model | Params | Classification | Retrieval | VQA | Grounding | **IND** | **OND** | **Overall Avg.** |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| Ours (SaHa-Qwen2-VL-2B) | 8.3B | 69.1 | 74.1 | 67.3 | 88.1 | **76.4** | **67.4** | **72.4** |
+
+
+## Citation
+
+If you find this model useful in your research, please cite our paper:
+
+```bibtex
+@misc{ju2025generatorembedder,
+      title={From Generator to Embedder: Harnessing Innate Abilities of Multimodal LLMs via Building Zero-Shot Discriminative Embedding Model}, 
+      author={Yeong-Joon Ju and Seong-Whan Lee},
+      year={2025},
+      eprint={2508.00955},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG},
+      url={https://arxiv.org/abs/2508.00955}, 
+}
+```

adapter_config.json

@@ -0,0 +1,43 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": {
+    "base_model_class": "Qwen2VLForConditionalGeneration",
+    "parent_library": "src.vlm_backbone.qwen2_vl.modeling_qwen2_vl"
+  },
+  "base_model_name_or_path": "Qwen/Qwen2-VL-7B-Instruct",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": "gaussian",
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.1,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "q_proj",
+    "gate_up_proj",
+    "v_proj",
+    "k_proj",
+    "down_proj",
+    "o_proj",
+    "qkv_proj",
+    "out_proj"
+  ],
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": true,
+  "use_rslora": false
+}

adapter_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ad4b04241fa2bb5b35831424237b521a8810eea085c566809cfa1cbb3bdf38c
+size 41841050

added_tokens.json

@@ -0,0 +1,16 @@
+{
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

chat_template.json

@@ -0,0 +1,3 @@
+{
+  "chat_template": "{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n{% endif %}<|im_start|>{{ message['role'] }}\n{% if message['content'] is string %}{{ message['content'] }}<|im_end|>\n{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>\n{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+}

config.json

@@ -0,0 +1,59 @@
+{
+  "_attn_implementation_autoset": true,
+  "architectures": [
+    "Qwen2VLForConditionalGeneration"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 3584,
+  "image_token_id": 151655,
+  "initializer_range": 0.02,
+  "intermediate_size": 18944,
+  "max_position_embeddings": 32768,
+  "max_window_layers": 28,
+  "model_type": "qwen2_vl",
+  "num_attention_heads": 28,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 4,
+  "padding_side": "left",
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": {
+    "mrope_section": [
+      16,
+      24,
+      24
+    ],
+    "rope_type": "default",
+    "type": "default"
+  },
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.0",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "video_token_id": 151656,
+  "vision_config": {
+    "depth": 32,
+    "embed_dim": 1280,
+    "hidden_act": "quick_gelu",
+    "hidden_size": 3584,
+    "in_channels": 3,
+    "in_chans": 3,
+    "mlp_ratio": 4,
+    "model_type": "qwen2_vl",
+    "num_heads": 16,
+    "patch_size": 14,
+    "spatial_merge_size": 2,
+    "spatial_patch_size": 14,
+    "temporal_patch_size": 2,
+    "torch_dtype": "bfloat16"
+  },
+  "vision_end_token_id": 151653,
+  "vision_start_token_id": 151652,
+  "vision_token_id": 151654,
+  "vocab_size": 152064
+}

preprocessor_config.json

@@ -0,0 +1,29 @@
+{
+  "do_convert_rgb": true,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.48145466,
+    0.4578275,
+    0.40821073
+  ],
+  "image_processor_type": "Qwen2VLImageProcessor",
+  "image_std": [
+    0.26862954,
+    0.26130258,
+    0.27577711
+  ],
+  "max_pixels": 12845056,
+  "merge_size": 2,
+  "min_pixels": 3136,
+  "patch_size": 14,
+  "processor_class": "Qwen2VLProcessor",
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "max_pixels": 12845056,
+    "min_pixels": 3136
+  },
+  "temporal_patch_size": 2
+}

special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "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|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:49a5b72e53f421c7c77ff9c9a0462a914dc50ca236e811c3bf13a06f77c5d799
+size 11420469

tokenizer_config.json

@@ -0,0 +1,145 @@
+{
+  "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
+    }
+  },
+  "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": "{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n{% endif %}<|im_start|>{{ message['role'] }}\n{% if message['content'] is string %}{{ message['content'] }}<|im_end|>\n{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>\n{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 32768,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "left",
+  "processor_class": "Qwen2VLProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}