This repository provides a merged model built from unsloth/Qwen2.5-7B-Instruct plus two separately trained LoRA adapters (DBBench / ALFWorld), combined with a weighted merge (alpha) and then merged into base weights. This repository contains merged model weights (not LoRA-only).

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.

Training Configuration

Base model: unsloth/Qwen2.5-7B-Instruct
Method: LoRA (trained separately for DBBench and ALFWorld)
LoRA: r=64, alpha=128

DB-LoRA (DB Bench)

Max sequence length: 2048
Epochs: 5
Learning rate: 2e-5
Dataset: u-10bei/dbbench_sft_dataset_react_v4 (edited local JSONL)

ALF-LoRA (ALF World)

Max sequence length: 2048
Epochs: 3
Learning rate: 2e-05
Dataset: u-10bei/sft_alfworld_trajectory_dataset_v5 (used as-is)

Adapter Merge (alpha)

Two LoRA adapters were trained independently:

DB-LoRA (DB Bench)
ALF-LoRA (ALF World)

They were combined using a weighted merge with coefficient alpha:

combined_adapter = DB + alpha * ALF
alpha was tuned empirically (e.g., alpha=0.6 used in the best run)

The combined adapter was then merged into the base model weights (merge_and_unload) and uploaded as a merged model.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = HF_UPLOAD_REPO_ID  # merged model repo

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

Sources & Terms (IMPORTANT)

Training data (ALFWorld): u-10bei/sft_alfworld_trajectory_dataset_v5 (used as-is). Training data (DBBench): u-10bei/dbbench_sft_dataset_react_v4 with additional preprocessing/edits (e.g., fixes for INSERT-style samples). The edited version was used for DB-LoRA training. Note: DB and ALF adapters were trained separately, then combined with alpha-weighted merge (see “Adapter Merge (alpha)”).

Dataset License: MIT License. This dataset is 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.

ikedabent/llm_ad_qwen2.5-7b-mix201

unsloth/Qwen2.5-7B-Instruct