SELEE/qwen3-4b-agent-full

Full Fine-Tuned Model: Qwen/Qwen3-4B-Instruct-2507

This repository provides full fine-tuned model weights (not a LoRA adapter) trained on agent trajectory datasets (ALFWorld + DBBench).

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("/content/full_sft_agentbench_qwen3_4b/checkpoint-600")
tokenizer = AutoTokenizer.from_pretrained("/content/full_sft_agentbench_qwen3_4b/checkpoint-600")

Training Details

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Training type: Full fine-tuning
• Datasets: ALFWorld trajectory dataset + DBBench SFT dataset
• Max sequence length: 4096
• Learning rate: 2e-06
• Epochs: 2