matsuo-llm-advanced-phase-f1a

Fine-tuned from Qwen/Qwen2.5-7B-Instruct for agent tasks.

Training Configuration

• LoRA: r=8, alpha=16 (Phase D identical)
• lr: 1e-5, epochs: 0.3, batch: 4×4=16
• Data: Spider/BIRD 70% + DBBench v4 20% (weak category weighted) + ALFWorld 10%, 3500 samples

Data Changes from Phase D

• Spider/BIRD: identical to Phase D (2450 random samples)
• ALFWorld v5: identical to Phase D (350 random samples)
• DBBench v4: resampled 700 from 1200 with weak category emphasis
  • Prioritized: INSERT, COUNT, AGG-MAX, AGG-SUM (weak in E6-a evaluation)
  • Category selection is rule-based (keyword matching), not LLM-based

Datasets

• u-10bei/dbbench_sft_dataset_react_v4 — Listed in the organizer-shared Phase B dataset list. Used as provided (no modification). Third-party synthetic SFT for DBBench format alignment; all tables, data, and queries are independently generated (per dataset description: "to avoid test data leakage").
• xlangai/spider — CC BY-SA 4.0 (Yale/Columbia Spider project)
• birdsql/bird_mini_dev — CC BY-SA 4.0 (HKU)
• Official Phase B ALFWorld v5 dataset — Organizer-provided, used as provided.

Compliance

• Evaluation data not used in training: No analysis of evaluation test data was conducted.
• LLM was not used for data quality filtering or selection.
• Category classification is rule-based (SQL keyword matching: INSERT, COUNT, MAX, SUM, etc.)
• Inference code not modified.

Usage

Compatible with vLLM v0.13.0+.