cite_sft

ScholarOwl SFT Selectors

This repository contains two PEFT/QLoRA adapters used as relevance judges in ScholarOwl, an academic-paper search and recommendation system.

The base-model weights are not included. Download them from their original publishers and observe the upstream licenses. Each subdirectory contains a sanitized adapter_config.json, an adapter_model.safetensors file and a model card.

Training provenance

• training source: PaSa sft_selector, 19,826 examples;
• method: QLoRA, one epoch;
• LoRA rank 16, alpha 32;
• hardware: one RTX 4090;
• measured Qwen3 adapter training time: approximately 2.1 hours.

Evaluation boundary

On the local 200-example Selector evaluation, Qwen3-SFT obtained Precision 0.93, Recall 0.79 and F1 0.851. On RealScholarQuery with a frozen retrieval result, Qwen3-SFT obtained Recall@20 0.3623 and final-set F1 0.3080; Qwen2.5-SFT obtained Recall@20 0.3460 and final-set F1 0.2633. The final dual-judge ScholarOwl pipeline obtained final-set F1 0.3258 under the fixed-corpus protocol.

These are different evaluation levels. Selector classification F1, fixed Top-k ranking F1 and thresholded final-set F1 must not be interchanged.

Code

ScholarOwl source code and reproducible result files:

https://github.com/Smife1023/cite

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3-8B"
tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(base, "lucas1023/cite_sft", subfolder="qwen3-sft")

Replace the base model and subfolder with Qwen/Qwen2.5-7B-Instruct and qwen2.5-sft for the second adapter.

License and notices

The adapters are incremental weights trained from third-party base models. Their use remains subject to the corresponding Qwen model licenses. PaSa data and paper metadata retain their original terms. The Apache-2.0 file in this repository does not relicense third-party base weights or datasets.