matsuo-llm-advanced-phase-x1

Fine-tuned from unsloth/Qwen2.5-7B-Instruct for agent tasks (DB + ALFWorld).

Architecture

Two-stage training with weighted LoRA merge:

1. Phase D reproduction (unsloth base): Same data/hyperparams as Phase D but using unsloth/Qwen2.5-7B-Instruct
2. DB-specific LoRA: Trained on 480 teacher-generated DB traces (Qwen3-32B-AWQ)
3. Weighted merge: final = phase_d_weight + 0.2 * (lora_merged - phase_d_weight)

Training Configuration

Stage 1: Phase D Reproduction

• Base: unsloth/Qwen2.5-7B-Instruct
• LoRA: r=8, alpha=16
• lr: 1e-5, epochs: 0.3, batch: 4x4=16
• Data: Spider/BIRD 70% + DBBench v4 20% + ALFWorld v5 10% (3500 samples)

Stage 2: DB LoRA

• Base: Stage 1 merged model
• LoRA: r=8, alpha=16
• lr: 5e-6, epochs: 0.25, batch: 2x8=16
• Data: 480 teacher-generated DB traces (Qwen3-32B-AWQ distilled)

Stage 3: Weighted Merge

• Lambda: 0.2 (conservative — preserves ALFWorld ability)
• Formula: final = base + λ * (full_merge - base)

Teacher Signal Generation

• Teacher model: Qwen/Qwen3-32B-AWQ (whitelisted for distillation/synthesis)
• Input: Spider/BIRD table schemas only (no AgentBench/ALFWorld data)
• 500 generated, 480 validated (regex + sqlparse filtering, no LLM)
• SQL type distribution: SELECT 73%, INSERT 18%, DELETE 5%, UPDATE 4%
• Weak category emphasis: INSERT, counting, agg-max, agg-sum

Datasets

• u-10bei/dbbench_sft_dataset_react_v4 — Listed in the organizer-shared Phase B dataset list. Used as provided (no modification).
• 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 DB traces generated by Qwen/Qwen3-32B-AWQ (whitelisted model)

Compliance

• Evaluation data not used in training.
• LLM was not used for data quality filtering or selection.
• Teacher signals generated from whitelisted model (Qwen/Qwen3-32B-AWQ) using Spider/BIRD schemas only.
• Filtering is regex/sqlparse only (non-LLM).
• Inference code not modified.
• Base model: unsloth/Qwen2.5-7B-Instruct (whitelisted learning-designated model).

Usage

Compatible with vLLM v0.13.0+.