qwen2.5-7b-instruct-sft-v6

This repository provides a merged full model produced by supervised fine-tuning for AgentBench-oriented ALFWorld/DBBench robustness.

Training Objective

Improve strict action selection reliability for ALFWorld prompts and strengthen SQL error-recovery robustness for DBBench prompts, while keeping balanced mixed-task behavior.

Training Configuration

• Method: SFT (Unsloth LoRA) + merge to full model
• Base model ID (upstream): Qwen/Qwen2.5-7B-Instruct
• Initialization model for this stage: prior merged checkpoint from the previous advanced retraining stage
• LoRA: r=16, alpha=32, dropout=0.0
• LoRA target modules: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
• Output learn mode: from_marker (ACTION: and Action: markers)
• Max sequence length: 4096
• Max steps: 400
• Epochs: 1
• Learning rate: 2.0e-6
• Per-device train batch size: 1
• Per-device eval batch size: 2
• Gradient accumulation steps: 32
• Effective global batch size: 32
• Warmup ratio: 0.03
• Weight decay: 0.01
• Eval/Save steps: 50 / 25

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "uchkw/qwen2.5-7b-instruct-sft-v6"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

Training Data / Sources & License (IMPORTANT)

• Primary source datasets:
  • u-10bei/sft_alfworld_trajectory_dataset_v5
  • u-10bei/dbbench_sft_dataset_react_v4
  • u-10bei/dbbench_sft_dataset_react_v3
• Data construction policy (concise):
  • ALFWorld samples were converted into strict one-line action supervision (ACTION: ...) with exact matching against AVAILABLE ACTIONS.
  • Added hard-copy style ALF augmentation to reinforce exact action copying and reduce formatting drift.
  • Mixed DBBench supervision and recovery-oriented examples for Unknown column style failures.
  • Mixed train ratio was controlled at approximately ALF:DB = 55:45.
• Dataset scale (fix8 stage2):
  • Train samples: 138496
  • Validation samples: 7289
  • Train ALF rows: 76173
  • Train DB rows: 62323
  • ALF strict-match in training set: 1.0
  • ALF completion-verb ratio: 0.4726
  • ALF toggle rows: 1772
• Evaluation snapshot (checkpoint-350, official_v02 setting):
  • DB overall_cat_accuracy: 0.4993505979
  • ALF success_rate: 0.60
  • ALF invalid_action_rate: 0.10
  • ALF task_limit_rate: 0.30
• Compliance:
  • Follow each source dataset card and license terms.
  • Follow base model terms of use.