This repository contains a LoRA adapter fine-tuned from unsloth/Qwen3.5-4B for a safe scientific research assistant task.
The model is designed to answer research-related questions using the provided context, extract supporting evidence, and clearly state limitations when the context is incomplete.
Model Description
This model is a fine-tuned LoRA adapter for scientific question answering and short research summarization.
The expected output format is:
text
1Answer:
2Evidence:
3Limitation:
The model is not intended to freely invent information. It should answer based on the given context and mention limitations when the context does not contain enough information.
Base Model
unsloth/Qwen3.5-4B
Fine-Tuning Method
This model was fine-tuned using:
LoRA
Unsloth
PyTorch
Transformers
PEFT
Manual training loop
The training setup used a conservative safe configuration to reduce unstable generation, repetition, and malformed output.
Datasets
The model was trained using scientific question-answering and summarization datasets:
Dataset
Usage
allenai/qasper
Scientific question answering with evidence
allenai/scitldr
Scientific TL;DR summarization
QASPER samples without evidence were filtered out so the model learns to provide grounded answers instead of unsupported evidence.
Intended Use
This model is intended for:
Scientific question answering
Research paper understanding
Context-based summarization
Evidence-based explanation
Educational research assistant experiments
Example prompt:
text
1Context:
2A research paper abstract or paragraph.
34Question:
5What is the main contribution of this paper?
Expected response:
text
1Answer:
2The paper proposes ...
34Evidence:
5The context states that ...
67Limitation:
8The provided context does not include ...
Not Intended For
This model should not be used for:
Medical diagnosis
Legal advice
Financial advice
Fully automated academic writing
Producing unsupported scientific claims
Answering without provided context
Replacing expert review
The model may still hallucinate, especially if the prompt does not provide enough context.
1prompt ="""
2You are a safe AI research assistant.
3Answer only using the given context.
45Context:
6Deep learning models have been used to improve scientific document understanding by extracting relevant evidence from papers.
78Question:
9What is the main idea of the context?
1011Respond using this format:
12Answer:
13Evidence:
14Limitation:
15"""1617inputs = tokenizer(prompt, return_tensors="pt").to(model.device)1819with torch.no_grad():20 outputs = model.generate(21**inputs,22 max_new_tokens=256,23 temperature=0.2,24 top_p=0.9,25 do_sample=False26)2728print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Recommended Prompt Format
Use this structure for best results:
text
1You are a safe AI research assistant.
2Answer only using the given context.
34Context:
5[Insert paper abstract, paragraph, or evidence here]
67Question:
8[Insert research question here]
910Respond using this format:
11Answer:
12Evidence:
13Limitation:
Output Format
The model should answer with:
text
1Answer:
2A direct answer to the question.
34Evidence:
5Relevant evidence from the given context.
67Limitation:
8A clear limitation if the context is incomplete.
Limitations
This model has several limitations:
It may hallucinate if the context is too short or unclear.
It is not a production-grade research assistant.
It does not verify facts from external sources.
It depends heavily on the quality of the provided context.
It may fail on highly technical papers outside the training distribution.
It may still produce weak or generic evidence if the prompt is poorly structured.
Safety Notes
The model was trained with a safer output structure, but it is still a language model. Outputs should be reviewed manually before being used in academic, scientific, or professional work.
Do not treat the model output as a verified citation or final academic conclusion.
Project Status
This is an experimental fine-tuned LoRA adapter for research assistant behavior. The goal is to improve grounded answering, evidence extraction, and limitation awareness in scientific QA tasks.
License
This model is released under the Apache License 2.0.
