ai-research-assistant-qwen35-4b-lora-safe

---

AI Research Assistant Qwen3.5-4B LoRA Safe

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:

Answer:
Evidence:
Limitation:

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:

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:

Context:
A research paper abstract or paragraph.

Question:
What is the main contribution of this paper?

Expected response:

Answer:
The paper proposes ...

Evidence:
The context states that ...

Limitation:
The 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.

Training Configuration

Main configuration used in the notebook:

MODEL_NAME = "unsloth/Qwen3.5-4B"
TRAINING_PROFILE = "safe"
MAX_SEQ_LENGTH = 2048
LOAD_IN_4BIT = False
QASPER_MAX_PAPERS = 500
SCITLDR_MAX_ROWS = 800
MAX_UPDATE_STEPS = 300
PER_DEVICE_BATCH_SIZE = 1
GRAD_ACCUM_STEPS = 8
LEARNING_RATE = 5e-5

The safe profile was used to reduce the risk of adapter collapse and unstable output.

How to Use

Install the required libraries:

pip install transformers peft accelerate bitsandbytes

Example loading code:

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

base_model = "unsloth/Qwen3.5-4B"
adapter_model = "YOUR_USERNAME/YOUR_REPO_NAME"

tokenizer = AutoTokenizer.from_pretrained(base_model)

model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.float16,
    device_map="auto"
)

model = PeftModel.from_pretrained(model, adapter_model)
model.eval()

Example inference:

prompt = """
You are a safe AI research assistant.
Answer only using the given context.

Context:
Deep learning models have been used to improve scientific document understanding by extracting relevant evidence from papers.

Question:
What is the main idea of the context?

Respond using this format:
Answer:
Evidence:
Limitation:
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.2,
        top_p=0.9,
        do_sample=False
    )

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Recommended Prompt Format

Use this structure for best results:

You are a safe AI research assistant.
Answer only using the given context.

Context:
[Insert paper abstract, paragraph, or evidence here]

Question:
[Insert research question here]

Respond using this format:
Answer:
Evidence:
Limitation:

Output Format

The model should answer with:

Answer:
A direct answer to the question.

Evidence:
Relevant evidence from the given context.

Limitation:
A 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.