paper-qa-lora

QLoRA adapter fine-tuned on Qwen/Qwen2.5-1.5B-Instruct to answer questions strictly from provided context: cite the source sentence, and say so plainly when the context doesn't contain the answer. Trained on QASPER-derived (context, question, answer) tuples, including explicit negative (unanswerable) examples. See the project plan for the full RAG + PEFT architecture this adapter is designed to slot into.

Training

Base model	Qwen/Qwen2.5-1.5B-Instruct
Dataset	whosouravsharma/paper-qa-qasper-sft
Method	QLoRA (4-bit NF4)
LoRA rank / alpha / dropout	16 / 32 / 0.05
Target modules	q_proj, k_proj, v_proj, o_proj
Trainable params	4,358,144 / 892,974,592 (0.488%)
Epochs	3.0
Learning rate	0.0002
Trained	2026-08-30T19:12:20.096276+00:00

Eval (final epoch, validation split)

Metric	Value
eval_loss	1.302024245262146
eval_runtime	197.317
eval_samples_per_second	8.692
eval_steps_per_second	2.174
eval_entropy	1.2851145122990464
eval_num_tokens	3205503.0
eval_mean_token_accuracy	0.7254449254149323
epoch	3.0

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "whosouravsharma/paper-qa-lora")
tokenizer = AutoTokenizer.from_pretrained("whosouravsharma/paper-qa-lora")

Limitations

Trained only on QASPER (NLP papers) — grounding/refusal behavior may not transfer as cleanly to other scientific domains. Evaluated on a held-out QASPER test slice (see the project's eval/ benchmark); no held-out custom-paper eval yet.