GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
AQ-MedAI
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
AQ-MedAI
Model Tree
Qwen/Qwen2.5-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
This is the model trained in our paper: GroupRank: A Groupwise Reranking Paradigm Driven by Reinforcement Learning (📝arXiv). Please refer our 🧩github repository for the usage of GroupRank-32B.
GroupRank is a reinforcement-learning-powered reranker that breaks the traditional “pointwise vs. listwise” trade-off: it feeds the query together with a small group of candidates to an LLM, lets the model perform within-group comparisons, and outputs individual relevance scores—combining the flexibility of pointwise scoring with the contextual awareness of listwise ranking. Training is driven by GRPO and a heterogeneous reward (NDCG + distributional alignment) that keeps scores consistent across groups, while a synthetic data pipeline eliminates the need for large human-labeled sets. On the reasoning-heavy BRIGHT and R2MED benchmarks, GroupRank sets (7B / 32B) new SOTA NDCG@10 (46.8 / 52.3) with strong generalization to classic retrieval tasks.
BRIGHT 基准测试结果 (NDCG@10)
| Models | Bio. | Earth. | Econ. | Psy. | Rob. | Stack. | Sus. | Leet. | Pony | AoPS | TheoQ. | TheoT. | Avg. |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Diver-Retriever-4B | 52.52 | 53.59 | 33.75 | 45.16 | 28.38 | 30.39 | 35.00 | 12.97 | 14.74 | 9.83 | 42.51 | 36.32 | 32.93 |
| Non-reasoning reranker | |||||||||||||
| rankT5 (3B) | 33.01 | 22.80 | 18.96 | 8.61 | 2.28 | 10.07 | 23.94 | 11.97 | 34.93 | 8.96 | 19.69 | 11.70 | 17.24 |
| RankZephyr (7B) | 42.60 | 19.51 | 18.79 | 29.85 | 13.98 | 13.33 | 29.29 | 12.97 | 31.37 | 7.44 | 31.19 | 31.67 | 23.50 |
| Reasoning reranker | |||||||||||||
| Rank-R1 (7B) | 40.85 | 31.38 | 23.17 | 32.04 | 19.82 | 11.23 | 35.53 | 3.77 | 6.60 | 4.99 | 17.20 | 30.10 | 21.39 |
| Rank-R1 (14B) | 49.57 | 41.15 | 27.56 | 40.04 | 28.46 | 28.19 | 43.76 | 6.96 | 18.45 | 7.88 | 34.80 | 43.11 | 30.83 |
| Rank-K (32B) | 51.06 | 42.34 | 32.95 | 44.52 | 33.07 | 28.44 | 41.57 | 12.82 | 21.62 | 8.48 | 39.31 | 43.15 | 33.28 |
| ReasonRank (7B)-w20s10 | 51.63 | 43.43 | 32.40 | 43.99 | 31.02 | 25.63 | 39.81 | 15.38 | 20.07 | 6.95 | 38.91 | 40.69 | 32.49 |
| ReasonRank (32B)-w20s10 | 53.89 | 47.59 | 36.34 | 52.64 | 36.48 | 34.17 | 44.47 | 15.21 | 14.81 | 5.47 | 40.63 | 45.29 | 35.58 |
| GroupRank-7B | 56.85 | 53.06 | 35.94 | 48.75 | 36.65 | 35.05 | 43.99 | 15.48 | 15.53 | 10.64 | 41.45 | 46.40 | 36.65 |
| GroupRank-32B | 59.48 | 56.49 | 40.12 | 50.46 | 38.36 | 39.16 | 43.32 | 19.48 | 16.32 | 13.34 | 46.39 | 47.91 | 39.24 |
python
from vllm import LLM, SamplingParamsfrom collections import defaultdictfrom transformers import AutoTokenizerimport randomimport torch random.seed(666) sys_prompt = '''Your task is to evaluate and rank documents based on how well they help answer the given query. Follow this evaluation priority:1. PRIMARY: Usefulness & Helpfulness - Does the document provide actionable information, solutions, or direct answers that help address the user's needs?2. SECONDARY: Relevance - Does the document contain information related to the query topic? Evaluation Process:1. First, identify the user's core intent and what kind of help they need from the query2. For each document, assess: - How directly it addresses the user's intent - What actionable information or answers it provides - How much it helps solve the user's problem or need3. Compare documents against each other to ensure proper ranking4. Assign scores that reflect the relative usefulness ranking Scoring Scale (0-10):- 9-10: Extremely helpful, directly answers the query with actionable information- 7-8: Very helpful, provides substantial useful information for the query- 5-6: Moderately helpful, contains some useful information but incomplete- 3-4: Minimally helpful, limited useful information despite topic relevance- 1-2: Barely helpful, mentions related topics but provides little useful information- 0: Not helpful at all, cannot assist with answering the query''' user_prompt = '''I will provide you {TOPK} documents, each indicated by a numerical identifier []. Score these documents based on their Usefulness and Relevance to the query.Query:{QUERY} Documents:{PASSAGES} ## Final Output FormatYou must structure your response in exactly two parts: provide your brief reasoning process first, then output final scores in JSON format like below, with document IDs as string keys and integer scores as values for all {TOPK} documents. The reasoning process and answer are enclosed within <reason> </reason> and <answer> </answer> tags, respectively. Do NOT output anything outside the specified tags. Follow this exact format:<reason> [Analyze each document's usefulness and relevance to the query, explaining your scoring rationale]</reason><answer> \```json{{"[1]": 5, "[2]": 3, "[3]": 8}}\``` </answer>''' class GroupReranker: def __init__(self, model_path, sys_prompt, user_prompt) -> None: # vllm offline inference self.llm = LLM(model=model_path, dtype="bfloat16", gpu_memory_utilization=0.9, tensor_parallel_size=torch.cuda.device_count(), max_model_len=32000) self.tokenizer = AutoTokenizer.from_pretrained(model_path) self.sampling_params = SamplingParams(temperature=0.3, top_p=0.8, max_tokens=8000, logprobs=10) self.group_system_prompt = sys_prompt self.group_user_prompt = user_prompt def rerank(self, query, doc_list): docs_str = ''.join(["[{}]. {}\n\n".format(idx+1, doc_text) for idx, doc_text in enumerate(doc_list)]) group_texts = self.group_user_prompt.format(QUERY=query, PASSAGES=docs_str, TOPK=len(doc_list)) message = self.tokenizer.apply_chat_template( [{'role': 'system', 'content': self.group_system_prompt}, {'role': 'user', 'content': group_texts}], tokenize=False, add_generation_prompt=True) output = self.llm.generate(message, self.sampling_params, use_tqdm=True) return output🌹 If you use this model, please ✨star our GitHub repository to support us. Your star means a lot!
from vllm import LLM, SamplingParamsfrom collections import defaultdictfrom transformers import AutoTokenizerimport randomimport torch random.seed(666) sys_prompt = '''Your task is to evaluate and rank documents based on how well they help answer the given query. Follow this evaluation priority:1. PRIMARY: Usefulness & Helpfulness - Does the document provide actionable information, solutions, or direct answers that help address the user's needs?2. SECONDARY: Relevance - Does the document contain information related to the query topic? Evaluation Process:1. First, identify the user's core intent and what kind of help they need from the query2. For each document, assess: - How directly it addresses the user's intent - What actionable information or answers it provides - How much it helps solve the user's problem or need3. Compare documents against each other to ensure proper ranking4. Assign scores that reflect the relative usefulness ranking Scoring Scale (0-10):- 9-10: Extremely helpful, directly answers the query with actionable information- 7-8: Very helpful, provides substantial useful information for the query- 5-6: Moderately helpful, contains some useful information but incomplete- 3-4: Minimally helpful, limited useful information despite topic relevance- 1-2: Barely helpful, mentions related topics but provides little useful information- 0: Not helpful at all, cannot assist with answering the query''' user_prompt = '''I will provide you {TOPK} documents, each indicated by a numerical identifier []. Score these documents based on their Usefulness and Relevance to the query.Query:{QUERY} Documents:{PASSAGES} ## Final Output FormatYou must structure your response in exactly two parts: provide your brief reasoning process first, then output final scores in JSON format like below, with document IDs as string keys and integer scores as values for all {TOPK} documents. The reasoning process and answer are enclosed within <reason> </reason> and <answer> </answer> tags, respectively. Do NOT output anything outside the specified tags. Follow this exact format:<reason> [Analyze each document's usefulness and relevance to the query, explaining your scoring rationale]</reason><answer> \```json{{"[1]": 5, "[2]": 3, "[3]": 8}}\``` </answer>''' class GroupReranker: def __init__(self, model_path, sys_prompt, user_prompt) -> None: # vllm offline inference self.llm = LLM(model=model_path, dtype="bfloat16", gpu_memory_utilization=0.9, tensor_parallel_size=torch.cuda.device_count(), max_model_len=32000) self.tokenizer = AutoTokenizer.from_pretrained(model_path) self.sampling_params = SamplingParams(temperature=0.3, top_p=0.8, max_tokens=8000, logprobs=10) self.group_system_prompt = sys_prompt self.group_user_prompt = user_prompt def rerank(self, query, doc_list): docs_str = ''.join(["[{}]. {}\n\n".format(idx+1, doc_text) for idx, doc_text in enumerate(doc_list)]) group_texts = self.group_user_prompt.format(QUERY=query, PASSAGES=docs_str, TOPK=len(doc_list)) message = self.tokenizer.apply_chat_template( [{'role': 'system', 'content': self.group_system_prompt}, {'role': 'user', 'content': group_texts}], tokenize=False, add_generation_prompt=True) output = self.llm.generate(message, self.sampling_params, use_tqdm=True) return output