qwen25_7b_lora_agentbench_v3

This repository provides a merged model fine-tuned from Qwen/Qwen2.5-7B-Instruct. The fine-tuning was performed using LoRA + Unsloth and the resulting adapter has been merged back into the base model weights.

This repository contains full model weights, making it ready for inference without the need to load a separate adapter.

Training Objective

This model is optimized for multi-turn agent tasks, specifically for ALFWorld (household navigation/interaction) and DBBench (database operations).

The training process applied loss to all assistant turns in the multi-turn trajectories, allowing the model to learn not just final answers, but also intermediate reasoning (Thought), environment observation processing, action selection, and error recovery.

Training Configuration

• Base model: Qwen/Qwen2.5-7B-Instruct
• Method: LoRA (merged post-training)
• Max sequence length: 2048
• Epochs: 2
• Learning rate: 2e-06
• LoRA Parameters: r=64, alpha=128

Usage

This model can be loaded using the standard transformers library or deployed with vLLM (recommended for evaluation).

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "your_hf_id/your_repo_name"

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