Base Model: Qwen/Qwen2.5-7B-Instruct
Training Method: Supervised Fine-Tuning (SFT) with LoRA
LoRA Configuration: r=32, alpha=32
Training Data: 50,000 synthetic samples generated by rule-based Python simulator

Training Details

Data Generation

Method: Pure rule-based Python simulator (no LLM)
Samples: 50,000 household task trajectories
Format: THOUGHT/ACTION dialogue format
Objects: 50 types (apple, book, mug, laptop, etc.)
Receptacles: 24 types (desk, drawer, shelf, fridge, etc.)
Rooms: 6 types (living room, kitchen, bedroom, etc.)

Training Configuration

Epochs: 2
Batch Size: 2 (effective: 32 with gradient accumulation)
Learning Rate: 2e-4
Optimizer: AdamW
Scheduler: Cosine
Final Loss: 0.09315

Quick Start

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "astom-M/matsuo-llm-advanced-household-agent"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")

# Example task
messages = [
    {"role": "user", "content": "You are in the middle of a room. Looking quickly around you, you see a desk 1, a drawer 1, a shelf 1.\n\nYour task is to: put a laptop on desk."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.7)
response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
print(response)

Expected Output Format

The model generates responses in THOUGHT/ACTION format:

THOUGHT: I need to find a laptop and put it on the desk.
ACTION: go to desk 1

Framework Versions

Transformers: 4.57.6
PyTorch: 2.10.0+cu128
PEFT: 0.11.0
Python: 3.10

License

Apache 2.0 (same as base model Qwen2.5-7B-Instruct)

Citation

bibtex

@misc{matsuo-llm-advanced-2025,
  author = {Matsuo Lab LLM Course 2025},
  title = {Household Task Agent - Advanced Competition},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/astom-M/matsuo-llm-advanced-household-agent}}
}

Model Development

This model was trained for the Matsuo Lab LLM Advanced Competition using:

Approved base model: Qwen2.5-7B-Instruct
Training method: SFT with LoRA (r=32, alpha=32)
Data: Rule-based synthetic household task trajectories (no ALFWorld or AgentBench data used)