Modelpublic

tomoniyukiwo/qwen25_7b_agentbench_lora

sourceHugging Faceapache-2.0updated 7mo agoView on Hugging Face

0likes5downloads

Model Card

qwen257bagentbench_lora

This repository provides a LoRA adapter fine-tuned from Qwen/Qwen2.5-7B-Instruct using LoRA + Unsloth + 3-Phase Training.

Note: This repository contains LoRA adapter weights only. The base model must be loaded separately.

Training Objective

This adapter is trained to improve multi-turn agent task performance on ALFWorld (household tasks) and DBBench (database operations).

Loss is applied to all assistant turns in the multi-turn trajectory, enabling the model to learn environment observation, action selection, tool use, and recovery from errors.

3-Phase Training Strategy

Phase	Data	Samples	Epochs	LR	Purpose
1	Original	3,702	3	2e-04	Foundation with high-quality data
2	Original + Synthetic	4,824	2	7e-05	Generalization via diverse scenarios
3	All (+ Distillation)	5,424	1	1e-06	Reasoning enhancement via distillation

Training Configuration

Parameter	Value
Base model	Qwen/Qwen2.5-7B-Instruct
Method	QLoRA (base + FP16 LoRA)
LoRA R / Alpha	64 / 128
RSLoRA	Enabled
Max seq length	8192
Optimizer	adamw_8bit
Gradient clip	1.0
LR scheduler	Cosine with warmup

Training Results

Phase	Final Train Loss	Time
Phase 1 (Foundation)	0.3189	1.2h
Phase 2 (Generalization)	0.0992	1.9h
Phase 3 (Reasoning)	0.0626	1.3h
Total		4.4h

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = "Qwen/Qwen2.5-7B-Instruct"
adapter = "tomoniyukiwo/qwen25_7b_agentbench_lora"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base,
    torch_dtype=torch.float16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter)

With Unsloth (faster)

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="tomoniyukiwo/qwen25_7b_agentbench_lora",
    max_seq_length=8192,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

Sources & Terms

—Training data: u-10bei/dbbench_sft_dataset_react_v4
—Dataset License: MIT License
—Base model: Qwen/Qwen2.5-7B-Instruct
—Compliance: Users must comply with the MIT license and the base model's original terms of use.