matsuo-llm-advanced-phase-f2a

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: 4x4=16
• Data: Spider/BIRD 2100 (60%) + Distilled 72B 500 (14%) + DBBench v4 700 (20%) + ALFWorld 200 (6%), 3500 samples

Data Composition

• Spider/BIRD: 2100 samples (60%) - reduced from Phase D to accommodate distilled data
• Distilled 72B: 500 samples (14%) - NEW: multi-turn DB conversations generated by Qwen2.5-72B-Instruct-AWQ
  • Generated from Spider/BIRD table schemas (not from AgentBench data)
  • Weak category emphasis (AGG-MAX, COUNT, INSERT, AGG-SUM)
  • Filtered by regex/sqlparse only (no LLM quality filtering)
• DBBench v4: 700 samples (20%) - same as Phase D
• ALFWorld v5: 200 samples (6%) - reduced from Phase D

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.
• Synthetic multi-turn data generated by Qwen2.5-72B-Instruct-AWQ (whitelisted model)

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.
• Distilled data generated from whitelisted model (Qwen2.5-72B-Instruct-AWQ) using Spider/BIRD schemas only.
• Filtering is regex/sqlparse only (non-LLM).
• Inference code not modified.

Usage

Compatible with vLLM v0.13.0+.