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/Qwen3-4B-Instruct-2507
Method: LoRA (full precision base)
Max sequence length: 2048
Epochs: 2
Learning rate: 5e-06
LoRA: r=128, alpha=128

Key Changes in 2c20

Based on 2c14 (best score: 4.8386)
- All successful patterns preserved
Added failure pattern prevention samples
- Prevent premature "Task succeeded" declaration
- Enforce inventory checking before actions
- Utilize AVAILABLE ACTIONS hints
Conservative upsampling for new samples
- Base ALF samples: 8×85=680
- Additional ALF samples: 3×15=45
- DB samples: 5×55=275
- Total custom: 1,000
All other parameters same as 2c14
- LR: 5.5e-6
- Epochs: 2
- DBBench 5% addition

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "TToyo2511/ttoyo_advance_2c20" #★TTT20260301 2c20版

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

Sources & Terms (IMPORTANT)

Training data:

u-10bei/sft_alfworld_trajectory_dataset_v5
u-10bei/dbbench_sft_dataset_react_v4 (5% sampled)

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.