Shin-YAM/Agent_try09

1---2base_model: Qwen/Qwen3-4B-Instruct-25073datasets:4- u-10bei/sft_alfworld_trajectory_dataset_v55- u-10bei/dbbench_sft_dataset_react_v46language:7- en8license: apache-2.09library_name: peft10pipeline_tag: text-generation11tags:12- lora13- agent14- tool-use15- alfworld16- dbbench17---18 19# ＜Qwen/Qwen3-4B-Instruct-2507/LoRA-combined_datasets--highLR--CleansedSQLdatasets-lessLORA＞20 21This repository provides a **LoRA adapter** fine-tuned from22**Qwen/Qwen3-4B-Instruct-2507** using **LoRA + Unsloth**.23 24This repository contains **LoRA adapter weights only**.25The base model must be loaded separately.26 27## Training Objective28 29This adapter is trained to improve **multi-turn agent task performance**30on ALFWorld (household tasks) and DBBench (database operations).31 32Loss is applied to **all assistant turns** in the multi-turn trajectory,33enabling the model to learn environment observation, action selection,34tool use, and recovery from errors.35 36The training process ic consist of two steps.37First, training for LoRA in order to be adapted to Database SQL,38Then Secondary, that for ALF is performed separately.39Finally, each LoRA adapter is merged into base model sequentially,40LoRA for DB and then that for ALF.41 42## Training Configuration43 44- Base model: Qwen/Qwen3-4B-Instruct-250745- Method: LoRA (full precision base)46- Max sequence length: 204847- Epochs: 248- Learning rate: 2e-0649- LoRA: r=64, alpha=12850 51## Usage52 53```python54from transformers import AutoModelForCausalLM, AutoTokenizer55from peft import PeftModel56import torch57 58base = "Qwen/Qwen3-4B-Instruct-2507"59adapter = "your_id/your-repo"60 61tokenizer = AutoTokenizer.from_pretrained(base)62model = AutoModelForCausalLM.from_pretrained(63    base,64    torch_dtype=torch.float16,65    device_map="auto",66)67model = PeftModel.from_pretrained(model, adapter)68```69 70## Sources & Terms (IMPORTANT)71 72Training data: u-10bei/dbbench_sft_dataset_react_v473 74Dataset License: MIT License. This dataset is used and distributed under the terms of the MIT License.75Compliance: Users must comply with the MIT license (including copyright notice) and the base model's original terms of use.76