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: 6e-06
LoRA: r=128, alpha=128

Key Changes in 2c17

Sample configuration: 100% maintained from 2c14
- ALF samples: 8 (same as 2c14)
- DB samples: 5 (same as 2c14)
- No sample addition (2c16 showed that adding samples degraded ALF)
Hyperparameter fine-tuning only
- LR: 5.5e-6 → 5.7e-6 (slight learning boost)
- ALF upsampling: 85x → 88x (ALF prioritization)
- DB upsampling: 55x → 52x
DBBench data ratio adjustment
- 5% → 4% (ALF stabilization)

Usage

python

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "TToyo2511/ttoyo_advance_2c17" #★TTT20260228 2c17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 (4% 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.