---
base_model: Qwen/Qwen2.5-7B-Instruct
datasets:
- u-10bei/sft_alfworld_trajectory_dataset_v5
- u-10bei/dbbench_sft_dataset_react_v4
- u-10bei/dbbench_sft_dataset_react_v3
language:
- en
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
- agentbench
- merged-model
- alfworld
- dbbench
- strict-action-copy
- unknown-column-recovery
- continual-sft
---

# qwen2.5-7b-instruct-sft-v5

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

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "uchkw/qwen2.5-7b-instruct-sft-v5"
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-50, official_v02 setting):
  - DB overall_cat_accuracy: `0.5180407064`
  - ALF success_rate: `0.62`
  - ALF invalid_action_rate: `0.14`
  - ALF task_limit_rate: `0.24`
- Compliance:
  - Follow each source dataset card and license terms.
  - Follow base model terms of use.