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 (A100 80GB Optimized)

Base model: Qwen/Qwen3-4B-Instruct-2507
GPU: A100 80GB (TF32 enabled, Flash Attention 2)
Method: LoRA (full precision base, bf16)
Datasets: ALFWorld v5 + DBBench v5 (combined ~5,500 samples)
Max sequence length: 8192
Epochs: 2
Learning rate: 1e-05
LoRA: r=256, alpha=512, dropout=0.05
Effective batch size: 8 x 2 = 16

Usage

python

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

base    = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "legoskier/qwen3-4b-agent-trajectory-lora_genspark_1"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="flash_attention_2",
)
model = PeftModel.from_pretrained(model, adapter)

Sources & Terms (IMPORTANT)

Training data:

u-10bei/sft_alfworld_trajectory_dataset_v5
u-10bei/sft_sql_dataset_trajectories_v5

Dataset License: MIT License. Users must comply with the MIT license and the base model's original terms of use.