qwen3-4b-agentbench-opd-merged-v2-sample

qwen3-4b-agentbench-opd-merged

This repository provides a merged model (LoRA weights merged into base) trained from Kaito-F/qwen3-4b-instruct-lora-v2 using OPD (On-Policy Distillation).

This is a standalone model - no adapter loading required.

Training Pipeline

1. Stage 1: SFT - Supervised fine-tuning on agent trajectory data
2. Stage 2: OPD - On-Policy Distillation from teacher model

Training Objective

This model is trained to improve multi-turn agent task performance on ALFWorld (household tasks) and DBBench (database operations) through knowledge distillation from a larger teacher model.

The OPD method reduces exposure bias by training on the student's own generations while guided by teacher model probabilities.

Training Configuration

Stage 2: OPD

• Student base: Kaito-F/qwen3-4b-instruct-lora-v2
• Teacher model: Qwen/Qwen3-30B-A3B-Instruct-2507
• Method: On-Policy Distillation with LoRA
• Max sequence length: 2048
• Steps: 100
• Learning rate: 1e-05
• LoRA: r=32, alpha=64

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Kaito-F/qwen3-4b-agentbench-opd-merged-v2-sample"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
)
# No PeftModel needed - weights are already merged!

Sources & Terms (IMPORTANT)

Training data:

• u-10bei/sft_alfworld_trajectory_dataset_v5
• u-10bei/dbbench_sft_dataset_react_v4

Dataset License: MIT License. This dataset is used and distributed under the terms of the MIT License. Compliance: Users must comply with the MIT license (including copyright notice) and the base model's original terms of use.