qwen3-4b-agent-sft-baseline-merged

baseline Merged Model

This repository provides a merged model derived from Qwen/Qwen3-4B-Instruct-2507 and a LoRA adapter trained with Unsloth.

Variant

• Name: baseline
• Purpose: Notebook baseline. Small effective batch and conservative LR.

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Dataset: u-10bei/sft_alfworld_trajectory_dataset_v5
• Method: LoRA + Unsloth SFT, then merged into the base model
• Max sequence length: 2048
• Epochs: 2
• Learning rate: 2e-06
• Gradient accumulation: 4
• LoRA: r=64, alpha=128, dropout=0.0
• Eval loss: 0.3616

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo_id = "magoemu/qwen3-4b-agent-sft-baseline-merged"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

Notes

• This repository contains merged full-model weights.
• Follow the base model license and the dataset terms when reusing this model.