qwen2.5-7b-instruct-opd-v1

This repository provides a merged full model produced by supervised fine-tuning for task-oriented instruction following.

Training Objective

Improve instruction following, action consistency, and response reliability in practical workflows.

Training Configuration

• Method: SFT (TRL SFTTrainer + Transformers, full-model)
• Base model ID: unsloth/Qwen2.5-7B-Instruct
• Validation ratio in checkpoint dataset build: 0.05
• Max sequence length: 512
• Max steps (actual): 4
• Epochs: 1
• Learning rate: 1e-6
• Per-device train batch size: 2
• Per-device eval batch size: 2
• Gradient accumulation steps: 8
• Effective global batch size: 16
• Warmup ratio: 0.1
• Weight decay: 0.05
• LR scheduler: cosine

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "uchkw/qwen2.5-7b-instruct-opd-v1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

Training Data / Sources & License (IMPORTANT)

• DBBench source dataset licenses:
  • u-10bei/dbbench_sft_dataset_react: MIT
  • u-10bei/dbbench_sft_dataset_react_v2: MIT
  • u-10bei/dbbench_sft_dataset_react_v3: MIT
  • u-10bei/dbbench_sft_dataset_react_v4: MIT
• DBBench data is further mixed with rule-based synthetic gap-fill samples, generated with:
  • teacher model: Qwen/Qwen2.5-72B-Instruct-AWQ
  • comparison student model: unsloth/Qwen2.5-7B-Instruct
• ALFWorld portion is rule-based synthetic worldtask data for instruction-following/action-format rehearsal.
• Compliance: Users must comply with the base model's terms of use.