--- license: apache-2.0 language: - en tags: - information-retrieval - LLM - Embedding - text-retrieval - disaster-management task_categories: - text-retrieval library_name: transformers dataset_tags: - DMIR01/DMRetriever_MTT --- This model is trained through the approach described in [DMRetriever: A Family of Models for Improved Text Retrieval in Disaster Management](https://www.arxiv.org/abs/2510.15087). The associated GitHub repository is available [here](https://github.com/KaiYin97/DMRETRIEVER). ## Model Overview **DMRetriever-335M** has the following features: - Model Type: Text Embedding - Supported Languages: English - Number of Paramaters: 335 - Context Length: 512 - Embedding Dimension: 1024 For more details, including model training, benchmark evaluation, and inference performance, please refer to our [paper](https://www.arxiv.org/abs/2510.15087), [GitHub](https://github.com/KaiYin97/DMRETRIEVER). ## DMRetriever series model list | **Model** | **Description** | **Backbone** | **Backbone Type** | **Hidden Size** | **#Layers** | |:--|:--|:--|:--|:--:|:--:| | [DMRetriever-33M](https://huggingface.co/DMIR01/DMRetriever-33M) | Base 33M variant | MiniLM | Encoder-only | 384 | 12 | | [DMRetriever-33M-PT](https://huggingface.co/DMIR01/DMRetriever-33M-PT) | Pre-trained version of 33M | MiniLM | Encoder-only | 384 | 12 | | [DMRetriever-109M](https://huggingface.co/DMIR01/DMRetriever-109M) | Base 109M variant | BERT-base-uncased | Encoder-only | 768 | 12 | | [DMRetriever-109M-PT](https://huggingface.co/DMIR01/DMRetriever-109M-PT) | Pre-trained version of 109M | BERT-base-uncased | Encoder-only | 768 | 12 | | [DMRetriever-335M](https://huggingface.co/DMIR01/DMRetriever-335M) | Base 335M variant | BERT-large-uncased-WWM | Encoder-only | 1024 | 24 | | [DMRetriever-335M-PT](https://huggingface.co/DMIR01/DMRetriever-335M-PT) | Pre-trained version of 335M | BERT-large-uncased-WWM | Encoder-only | 1024 | 24 | | [DMRetriever-596M](https://huggingface.co/DMIR01/DMRetriever-596M) | Base 596M variant | Qwen3-0.6B | Decoder-only | 1024 | 28 | | [DMRetriever-596M-PT](https://huggingface.co/DMIR01/DMRetriever-596M-PT) | Pre-trained version of 596M | Qwen3-0.6B | Decoder-only | 1024 | 28 | | [DMRetriever-4B](https://huggingface.co/DMIR01/DMRetriever-4B) | Base 4B variant | Qwen3-4B | Decoder-only | 2560 | 36 | | [DMRetriever-4B-PT](https://huggingface.co/DMIR01/DMRetriever-4B-PT) | Pre-trained version of 4B | Qwen3-4B | Decoder-only | 2560 | 36 | | [DMRetriever-7.6B](https://huggingface.co/DMIR01/DMRetriever-7.6B) | Base 7.6B variant | Qwen3-8B | Decoder-only | 4096 | 36 | | [DMRetriever-7.6B-PT](https://huggingface.co/DMIR01/DMRetriever-7.6B-PT) | Pre-trained version of 7.6B | Qwen3-8B | Decoder-only | 4096 | 36 | ## Usage Using HuggingFace Transformers: ```python import numpy as np import torch import torch.nn.functional as F from transformers import AutoTokenizer, AutoModel MODEL_NAME = "DMIR01/DMRetriever-335M" # Load model/tokenizer device = "cuda" if torch.cuda.is_available() else "cpu" dtype = torch.float16 if device == "cuda" else torch.float32 tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True) # Some decoder-only models have no pad token; fall back to EOS if needed if tokenizer.pad_token is None and tokenizer.eos_token is not None: tokenizer.pad_token = tokenizer.eos_token model = AutoModel.from_pretrained(MODEL_NAME, torch_dtype=dtype).to(device) model.eval() # Mean pooling over valid tokens (mask==1) def mean_pool(last_hidden_state: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor: mask = attention_mask.unsqueeze(-1).type_as(last_hidden_state) # [B, T, 1] summed = (last_hidden_state * mask).sum(dim=1) # [B, H] counts = mask.sum(dim=1).clamp(min=1e-9) # [B, 1] return summed / counts # [B, H] # Optional task prefixes (use for queries; keep corpus plain) TASK2PREFIX = { "FactCheck": "Given the claim, retrieve most relevant document that supports or refutes the claim", "NLI": "Given the premise, retrieve most relevant hypothesis that is entailed by the premise", "QA": "Given the question, retrieve most relevant passage that best answers the question", "QAdoc": "Given the question, retrieve the most relevant document that answers the question", "STS": "Given the sentence, retrieve the sentence with the same meaning", "Twitter": "Given the user query, retrieve the most relevant Twitter text that meets the request", } def with_prefix(task: str, text: str) -> str: p = TASK2PREFIX.get(task, "") return f"{p}: {text}" if p else text # Batch encode with L2 normalization (recommended for cosine/inner-product search) @torch.inference_mode() def encode_texts(texts, batch_size: int = 32, max_length: int = 512, normalize: bool = True): all_embs = [] for i in range(0, len(texts), batch_size): batch = texts[i:i + batch_size] toks = tokenizer( batch, padding=True, truncation=True, max_length=max_length, return_tensors="pt", ) toks = {k: v.to(device) for k, v in toks.items()} out = model(**toks, return_dict=True) emb = mean_pool(out.last_hidden_state, toks["attention_mask"]) if normalize: emb = F.normalize(emb, p=2, dim=1) all_embs.append(emb.cpu().numpy()) return np.vstack(all_embs) if all_embs else np.empty((0, model.config.hidden_size), dtype=np.float32) # ---- Example: plain sentences ---- sentences = [ "A cat sits on the mat.", "The feline is resting on the rug.", "Quantum mechanics studies matter and light.", ] embs = encode_texts(sentences) # shape: [N, hidden_size] print("Embeddings shape:", embs.shape) # Cosine similarity (embeddings are L2-normalized) sims = embs @ embs.T print("Cosine similarity matrix:\n", np.round(sims, 3)) # ---- Example: query with task prefix (QA) ---- qa_queries = [ with_prefix("QA", "Who wrote 'Pride and Prejudice'?"), with_prefix("QA", "What is the capital of Japan?"), ] qa_embs = encode_texts(qa_queries) print("QA Embeddings shape:", qa_embs.shape) ``` ## Citation If you find this repository helpful, please kindly consider citing the corresponding paper. Thanks! ``` @article{yin2025dmretriever, title={DMRetriever: A Family of Models for Improved Text Retrieval in Disaster Management}, author={Yin, Kai and Dong, Xiangjue and Liu, Chengkai and Lin, Allen and Shi, Lingfeng and Mostafavi, Ali and Caverlee, James}, journal={arXiv preprint arXiv:2510.15087}, year={2025} } ```