takayosh/agentbenchfinal

qwen2.5-7b-alf-dbb-merged-final

This repository provides a merged full model based on Qwen/Qwen2.5-7B-Instruct.

Model Construction Pipeline

1. 1.Train LoRA adapter on ALFWorld
2. 2.Train LoRA adapter on DBBench
3. 3.Merge adapters using ties (density=0.1)
4. 4.Apply additional stabilization fine-tuning (LoRA)
5. 5.Merge final adapter into base model

This repository contains full merged weights (no adapter required).

Final Training Configuration

• —Base model: Qwen/Qwen2.5-7B-Instruct
• —Merge method: ties
• —Merge density: 0.1
• —Final stage epochs: 1
• —Learning rate: 1e-05
• —Final LoRA: r=16, alpha=16
• —Max sequence length: 2024

Datasets

• —u-10bei/sftalfworldtrajectorydatasetv5
• —u-10bei/dbbenchsftdatasetreactv4

Additional distilled datasets were optionally included.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "takayosh/agentbenchfinal"

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

Sources & Terms (IMPORTANT)

Training data:

• —u-10bei/sftalfworldtrajectorydatasetv5
• —u-10bei/dbbenchsftdatasetreactv4

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.