Downloads · 30 days
0
vamsee9201/qwen35-9b-hybrid-rag-lora
qwen35-9b-hybrid-rag-lora is a text generation model from vamsee9201. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as apache-2.0.
This repository contains the Qwen3.5 9B LoRA adapter trained for the second Lean RAG experiment. It also hosts the sealed benchmark, retrieval measurements, answer-quality tables, paired statistical analysis, cost dat…
Downloads · 30 days
0
Access
Public
Updated Sep 7, 2026
Repo size
433 MB
Likes
0
Public
Click a slice to open those files.
.safetensors433 MB · 99%
From the Hugging Face model README
This repository contains the Qwen3.5 9B LoRA adapter trained for the second Lean RAG experiment. It also hosts the sealed benchmark, retrieval measurements, answer-quality tables, paired statistical analysis, cost data, and blinded human-review sheets.
The adapter was trained to answer questions from supplied document passages, cite document and page identifiers, and abstain when the evidence is insufficient. It was evaluated on public GovInfo documents.
This is a research adapter and reproducibility artifact. Validate it on your own corpus before relying on it in a production or high-stakes system.
Qwen/Qwen3.5-9Brun1/v0-20260907-054214/checkpoint-564| Generator | BM25 primary score | Vertex hybrid primary score | Hybrid exact citation recall |
|---|---|---|---|
| Gemini 3.8 Flash | 75.0% | 82.0% | 73% |
| Base Qwen3.5 9B | 76.5% | 77.0% | 57% |
| BM25-trained Qwen3.5 9B | 76.0% | 79.5% | 73% |
| This hybrid-trained adapter | 75.5% | 77.5% | 74% |
The hybrid-trained adapter completed successfully and produced the strongest exact citation recall. It did not outperform the earlier BM25-trained adapter on the primary answer score. The earlier adapter transferred to hybrid contexts without additional training and remains the recommended local configuration.
The hybrid-trained adapter scored 4.5 points below Gemini under hybrid retrieval. Its one-sided 95% lower confidence bound was -9 points, so it did not meet the predefined five-point non-inferiority criterion.
Download the selected adapter checkpoint:
hf download vamsee9201/qwen35-9b-hybrid-rag-lora \
--include 'run1/v0-20260907-054214/checkpoint-564/**' \
--local-dir qwen35-hybrid-rag
Run it with ms-swift:
swift infer \
--adapters qwen35-hybrid-rag/run1/v0-20260907-054214/checkpoint-564 \
--use_hf true \
--infer_backend transformers \
--max_length 8192 \
--max_new_tokens 220 \
--temperature 0 \
--enable_thinking false
The model expects a prompt containing a question and labeled evidence passages. The complete prompt construction code is available in the Lean RAG repository.
The evaluation/vertex-hybrid/ directory contains:
gemini-embedding-001.The Qwen3.5 9B base model is released under Apache 2.0. This adapter uses the same license. Review the base model license and terms before deployment.