This repository provides a merged model derived from GRPO-trained checkpoints based on mssfj/Qwen2.5-7B-Instruct_grpo_alfworld_trajectory_dataset and mssfj/Qwen2.5-7B-Instruct_dbbench_grpo_dataset_react.

The source GRPO models were trained with LoRA + Unsloth/TRL GRPO and then merged. This repository contains full merged model weights (not adapter-only weights).

Training Objective

This model 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.

Training Configuration

Base model: Qwen/Qwen2.5-7B-Instruct
Intermediate model (ALFWorld GRPO run): mssfj/Qwen2.5-7B-Instruct_sft_alfworld_trajectory_dataset_v5-13
Intermediate model (DBBench GRPO run): mssfj/Qwen2.5-7B-Instruct_dbbench_sft_dataset_react_v4-3
Merge base model: mssfj/Qwen2.5-7B-Instruct_grpo_alfworld_trajectory_dataset
Merge method: linear
Merge weights: 0.5 (ALFWorld GRPO) + 0.5 (DBBench GRPO)
Merge dtype: bfloat16
Method: LoRA (full precision base) + GRPO, then merge
Max sequence length: 2048
Epochs: N/A (step-based GRPO training)
Learning rate: 5e-5
LoRA: r=16, alpha=32
GRPO steps: 200 (ALFWorld), 125 (DBBench)
GRPO batch config: per_device_train_batch_size=4, gradient_accumulation_steps=2, num_generations=8

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "mssfj/Qwen2.5-7B-Instruct_alfworld_dbbench_grpo_merge"

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

Sources & Terms (IMPORTANT)

Training data:

u-10bei/sft_alfworld_trajectory_dataset
u-10bei/dbbench_sft_dataset_react

Dataset License: MIT License. These datasets are 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.