Instructions to use PaSaMaster/Ranker with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use PaSaMaster/Ranker with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="PaSaMaster/Ranker")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("PaSaMaster/Ranker")
model = AutoModelForCausalLM.from_pretrained("PaSaMaster/Ranker", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use PaSaMaster/Ranker with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "PaSaMaster/Ranker"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "PaSaMaster/Ranker",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/PaSaMaster/Ranker

SGLang

How to use PaSaMaster/Ranker with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "PaSaMaster/Ranker" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "PaSaMaster/Ranker",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "PaSaMaster/Ranker" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "PaSaMaster/Ranker",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use PaSaMaster/Ranker with Docker Model Runner:
```
docker model run hf.co/PaSaMaster/Ranker
```

Ranker / README.md

yuwendu

Add baseline comparisons to PaSaMaster results

6c595b0 verified about 1 month ago

preview code

Raw History Blame Contribute Delete

2.83 kB

	---
	license: apache-2.0
	base_model: Qwen/Qwen3-30B-A3B
	library_name: transformers
	pipeline_tag: text-generation
	tags:
	- pasamaster
	- reranker
	- scientific-literature-retrieval
	- evidence-grounded
	- qwen3
	---

	# PaSaMaster Ranker

	PaSaMaster Ranker is the evidence-grounded paper-ranking model used in PaSaMaster, a self-evolving agentic system for scientific literature retrieval. Built on Qwen3-30B-A3B, it is PaSaMaster's verification and ranking component—not a standalone search engine or general-purpose chat model.

	## How it works

	Given a research intent, a query-specific relevance checklist, and a candidate paper's metadata, abstract, and retrieved evidence, the Ranker:

	- scores each checklist criterion from 1 to 5;
	- provides evidence-grounded rationales and flags weak matches;
	- estimates holistic relevance and reranks verified papers.

	This design ranks authentic, traceable paper records instead of generating citations. The model was trained through multidisciplinary knowledge distillation on 42,762 query–paper pairs spanning 19 disciplines and 97 fine-grained topics, including positive, partial-match, and hard-negative examples.

	## PaSaMaster results

	The following are end-to-end PaSaMaster system results using this Ranker, measured on PaSaMaster-Bench (244 expert-curated tasks across 38 disciplines):

	\| Method \| NDCG@20 \| Recall@20 \| Precision@20 \| F1@20 \| Hallucination \| Cost/query \|
	\|---\|---:\|---:\|---:\|---:\|---:\|---:\|
	\| Google Scholar \| 2.07 \| 1.69 \| 1.48 \| 1.39 \| 0% \| — \|
	\| OpenScholar \| 14.61 \| 11.68 \| 8.52 \| 7.92 \| 0% \| — \|
	\| Bohrium Science Navigator \| 22.39 \| 19.37 \| 12.50 \| 12.26 \| 0% \| — \|
	\| DeepSeek-v3.2 \| 35.82 \| 24.76 \| 15.35 \| 15.56 \| 12.94% \| $0.28 \|
	\| Kimi-K2.5 \| 37.80 \| 28.08 \| 16.95 \| 17.36 \| 26.59% \| $0.16 \|
	\| MiniMax-M2.7 \| 30.70 \| 24.23 \| 14.42 \| 15.11 \| 32.66% \| $0.18 \|
	\| GLM-5 \| 35.89 \| 28.99 \| 16.93 \| 18.18 \| 21.64% \| $0.56 \|
	\| Gemini-3.1-pro \| 31.34 \| 21.30 \| 11.68 \| 12.48 \| 27.54% \| $0.38 \|
	\| GPT-5.2 \| 31.59 \| 25.32 \| 16.82 \| 16.69 \| 5.65% \| $6.06 \|
	\| Google Scholar Labs \| 30.54 \| 29.01 \| 18.79 \| 18.87 \| 0% \| — \|
	\| PaSaMaster \| 39.52 \| 33.24 \| 23.46 \| 23.00 \| 0% \| $0.05 \|

	PaSaMaster achieves 16.5× the F1@20 of Google Scholar and 37.8% higher F1@20 than GPT-5.2 in the reported evaluation, at approximately $0.05 per query.

	## Intended use

	Use this checkpoint within the PaSaMaster pipeline or a compatible evidence-grounded literature-ranking workflow. Supply retrieved paper evidence and treat its outputs as relevance judgments, not as independent proof that a citation is valid.

	## Resources

	- [Paper](https://arxiv.org/abs/2605.14306)
	- [Code](https://github.com/sjtu-sai-agents/PaSaMaster)
	- [PaSaMaster-Bench](https://huggingface.co/datasets/PaSaMaster/PaSaMaster_bench)