allenai/AstaBrief_8B_SFT

AstaBrief-8B-SFT

AstaBrief-8B-SFT is the intermediate supervised fine-tuning (SFT) checkpoint of AstaBrief-8B, a model designed to turn a research question and retrieved scientific literature excerpts into a cited report.

The model is initialized from Qwen3-8B and fine-tuned on AstaBrief_SFT_Mix, a collection of real user queries paired with reports generated from the multi-step Asta ScholarQA report generation pipeline, using a variety of backing models: Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini, and GPT-4.1.

For a detailed overview of the project, see our blog.

Inference and Usage

[!NOTE] Recommended prompt: This checkpoint was fine-tuned using this SFT prompt. For best results, we recommend using the same prompt format at inference time with your input query and section references. Using a different prompt or interaction format may lead to degraded or inconsistent behavior.

Following is an example inference code snippet:

from transformers import AutoModelForCausalLM, AutoTokenizer
from vllm import LLM, SamplingParams

model_name = "allenai/AstaBrief_8B_SFT"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input by providing your query and retrieved literature excerpts in our recommended prompt format
messages = [
    {"role": "user", "content": formatted_sft_prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

# instantiate sampling parameters
llm = LLM(model=model_name)
sampling_params = SamplingParams(
    temperature=0.7,
    top_p=0.95,
    max_tokens=4096,
    stop_token_ids=[tokenizer.eos_token_id],
)

# conduct text completion
outputs = llm(text, sampling_params)
output_text = outputs[0].outputs[0].text.strip()
output_ids = len(outputs[0].outputs[0].token_ids)

print("content:", output_text)

Evaluation Results

Our SFT training procedure improves the performance of the base Qwen3-8B model on the ScholarQA-CS2 test set, a set of 100 user-written computer science research questions.

Intended Uses and Limitations

This model is licensed under Apache 2.0 and is based on Qwen 3-8B. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines. Please refer to the details in our SFT dataset for more information about the data sources used in SFT training.

Training

SFT training was conducted in open-instruct on 8xH100s. Hyperparameter settings:

Links

📝 Project blog: https://allenai.org/blog/astabrief

🤗 SFT dataset: AstaBrief_SFT_Mix

🤖 Base model: Qwen3-8B

🤖 DPO checkpoint: AstaBrief_8B