YongdongWang/llama-3.2-1b-lora-qlora-dart-llm

Llama 3.2 1B - DART LLM Robot Task Planning (QLoRA Fine-tuned)

This model is a QLoRA fine-tuned version of meta-llama/Llama-3.2-1B specialized for robot task planning in construction environments.

The model converts natural language commands into structured task sequences for construction robots including excavators and dump trucks.

Model Details

• Base Model: meta-llama/Llama-3.2-1B
• Fine-tuning Method: QLoRA (4-bit quantization + LoRA)
• LoRA Rank: 16-32 (optimized per model size)
• LoRA Alpha: 16-32
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Dataset: YongdongWang/dart_llm_tasks_pretty
• Training Domain: Construction robotics

Usage

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

# Load tokenizer and base model
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.2-1B")
base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.2-1B",
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "YongdongWang/llama-3.2-1b-lora-qlora-dart-llm")

# Generate robot task sequence
instruction = "Deploy Excavator 1 to Soil Area 1 for excavation"
prompt = f"### Instruction:\n{instruction}\n\n### Response:\n"
inputs = tokenizer(prompt, return_tensors="pt")

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id
    )

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Training Details

• Training Data: DART LLM Tasks - Robot command and task planning dataset
• Domain: Construction robotics (excavators, dump trucks, soil/rock areas)
• Training Epochs: 6-12 (optimized per model size)
• Batch Size: 1 (with gradient accumulation)
• Learning Rate: 1e-4 to 3e-4 (scaled by model size)
• Optimizer: paged_adamw_8bit or adamw_torch

Capabilities

• Multi-robot coordination: Handle multiple excavators and dump trucks
• Task dependencies: Generate proper task sequences with dependencies
• Spatial reasoning: Understand soil areas, rock areas, puddles, and navigation
• Action planning: Convert commands to structured JSON task definitions

Example Output

The model generates structured task sequences in JSON format for robot execution:

{
  "tasks": [
    {
      "instruction_function": {
        "dependencies": [],
        "name": "target_area_for_specific_robots",
        "object_keywords": ["soil_area_1"],
        "robot_ids": ["robot_excavator_01"],
        "robot_type": null
      },
      "task": "target_area_for_specific_robots_1"
    }
  ]
}

Limitations

This model is specifically trained for construction robotics scenarios and may not generalize to other domains without additional fine-tuning.

Citation

@misc{llama_3.2_1b_lora_qlora_dart_llm,
  title={Llama 3.2 1B Fine-tuned with QLoRA for DART LLM Tasks},
  author={YongdongWang},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/YongdongWang/llama-3.2-1b-lora-qlora-dart-llm}
}

Model Card Authors

YongdongWang

Model Card Contact

For questions or issues, please open an issue in the repository or contact the model author.