shiwano/qwen2.5-7b-agent-sft-v13

Qwen2.5-7B-Instruct — Agent SFT

Fine-tuned model for AgentBench tasks (DB Bench + ALFWorld).

Base Model

• Model: Qwen/Qwen2.5-7B-Instruct
• Max sequence length (training): 3584

Training Configuration

• Method: SFT with High-Rank LoRA (merged)
• LoRA rank: 128
• LoRA alpha: 256
• LoRA dropout: 0
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Learning rate: 2e-05
• Epochs: 2
• Batch size: 1 (x8 gradient accumulation)
• Warmup ratio: 0.1
• LR scheduler: cosine
• Weight decay: 0.05
• Max grad norm: 1.0
• Precision: bf16

Training Data

• data/processed/db_bench_raw.jsonl
• data/processed/alfworld_augmented.jsonl
• data/synthetic/db_bench/trajectories-ranking-other.jsonl

Usage

This model can be served with vLLM:

vllm serve <model_path> --max-model-len 8192 --gpu-memory-utilization 0.95

License

Apache-2.0 (following the base model license)