GRASP checkpoints
Checkpoints, text-branch rankings and SFT data of GRASP (anonymous submission): plan-guided graph retrieval with adaptive fusion and reranking on the three STaRK semi-structured knowledge bases (Amazon, MAG, Prime). Code, setup and the exact commands are in the anonymous code repository: https://anonymous.4open.science/r/GRASP-6B75
The base models are not included; download them from Hugging Face:
Qwen/Qwen3-1.7B, Qwen/Qwen3-Embedding-4B, Qwen/Qwen3-Reranker-0.6B.
| Folder | What it is |
|---|---|
planner/<ds>/rft2_lora/ |
The planner's round-2 RFT LoRA (r=64, alpha=128) on Qwen/Qwen3-1.7B. Merge it into the base model (swift export --model Qwen/Qwen3-1.7B --adapters planner/<ds>/rft2_lora --merge_lora true, ms-swift 4.2.0, on a GPU). |
planner/<ds>/grpo2_lora/ |
The selected round-2 GRPO LoRA (r=32, alpha=64), trained on top of the merged RFT model (its base_model_name_or_path refers to that merged model, rft2_merged). vLLM serves the merged model with this LoRA (--enable-lora --max-lora-rank 64): together they are the final planner. |
reranker/<ds>/checkpoint-<step>/ |
The selected reranker LoRA (r=32, alpha=64) on Qwen/Qwen3-Reranker-0.6B, with grasp_reranker.json (serializer and instruction identity). |
sft/<ds>/sft.jsonl |
The planner's SFT set: 500 teacher plans (Qwen3.6-27B) for training queries, as chat rows (messages: the system prompt, the user message with the query and its entity table, the plan as the assistant turn). Training with N = 125 / 250 uses its first rows, and the ablation without the entity table uses the same rows with the query alone as the user message (the code repository's scripts/train/train_planner.sh does both). |
text/<ds>/{train,val,test}.qres |
The text-branch rankings of the train, validation and test queries: mFAR (Multi-Field Adaptive Retrieval, Li et al., ICLR 2025), top 100 per query, TREC format (tab-separated qid 0 doc_id 0 score 0). The code repository's scripts/infer/download_checkpoints.sh downloads them and installs them where the pipeline reads them (work/text/<ds>/<split>.qres). |
<ds> is amazon, mag or prime. Fusion settings used with these checkpoints (also in configs/released.json of the
code repository):
| Dataset | graph-text RRF (k, w_graph) | reranker RRF at K = 50 (k, w_reranker) | reranker step |
|---|---|---|---|
| Amazon | (3, 0.35) | (100, 0.90) | 913 |
| MAG | (30, 0.95) | (20, 0.70) | 737 |
| Prime | (20, 0.50) | (5, 0.70) | 1146 |
The merged planner is saved by ms-swift 4.2.0 in the transformers 5.x format (rope_parameters in config.json);
serve it with vLLM 0.21.0 / transformers >= 5.8.
License: the Qwen-derived weights follow the Apache-2.0 license of their base models.