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.

Data Processing

Train/Validation Split: 95% / 5%
Random Seed: 3407 (used for shuffling and initialization)
Loss Masking: Loss was computed only on the assistant's responses. User prompts and observations were masked during training (train_on_responses_only was applied to <|im_start|>assistant\n).

Training Configuration

Base model: Qwen/Qwen3-4B-Instruct-2507
Method: LoRA + Unsloth (Merged in 16-bit)
Max sequence length: 8192
Epochs: 1
Learning rate: 5e-06
LoRA: r=8, alpha=16
PER_DEVICE_TRAIN_BATCH_SIZE = 4
GRAD_ACCUM = 2
WARMUP_RATIO = 0.1
WEIGHT_DECAY = 0.05
NEFTUNE_NOISE_ALPHA = 5.0
VAL_RATIO = 0.05

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "choco800/qwen3-4b-agent-v22"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

Sources & Terms (IMPORTANT)

Training data:

u-10bei/sft_alfworld_trajectory_dataset_v5 (available on Hugging Face Hub)
u-10bei/sft_alfworld_trajectory_dataset_v4 (available on Hugging Face Hub)
u-10bei/sft_alfworld_trajectory_dataset_v3 (available on Hugging Face Hub)

Dataset License: MIT License. These datasets are used and distributed under the terms of the MIT License. Compliance: Users must comply with the dataset licenses and the base model's original terms of use (Apache 2.0).