---
language:
- my
- en
license: apache-2.0
tags:
- functiongemma
- function-calling
- tool-use
- burmese
- myanmar
- on-device
- gemma
base_model: google/functiongemma-270m-it
datasets:
- thantzinphyo/Hey-Function-Burmese
pipeline_tag: text-generation
---

<div align="center">
  <img src="hey_avatar.png" width="108" alt="Hey-Function-270M Avatar">
  <h1>Hey-Function-270M</h1>
  <p><strong>A Lightweight On-Device Burmese & English Function-Calling Agent</strong></p>
</div>

---

## Overview

Hey-Function-270M Alpha Version is an on-device function-calling agent designed for lightweight platforms (smartphones, IoT, edge boards). It enables applications to map Burmese, English, and Code-Switching natural language instructions into structured tool calls.

> [!NOTE]
> **Release:** This checkpoint serves as an initial proof-of-concept to evaluate Burmese function-calling feasibility on a compact 270M-parameter architecture.

- **Base Model:** Google/FunctionGemma-270M-IT
- **Dataset:** [thantzinphyo/Hey-Function-Burmese](https://huggingface.co/datasets/thantzinphyo/Hey-Function-Burmese)
- **Training Stage:** Initial Baseline (1 Epoch)

---

## Weight Adaptation & Training Methodology

How the model weights were updated and retrained:

- **Full-Parameter Tuning:**
- All 270M transformer decoder weights were unfrozen and updated without parameter-efficient adapters (such as LoRA).
- This was done because the original base model's tool-calling capability was trained almost exclusively in English.
- Full adaptation allowed attention representations to align with Myanmar Subject-Object-Verb (SOV) sentence structures and entity schemas.
- **Dynamic Candidate Collation:**
- Paired target ground-truth tools with 2 negative distractors per sample to prevent 61-tool schema overflow.
- Fitted comfortably within an extended 2,800-token context budget.
- Achieved 0.0% sequence truncation across the entire 42.6K dataset.
- **Response-Only Loss Masking:**
- Masked all system prompt, tool schema, and user query tokens with label -100.
- Gradients were computed strictly on the model's generated function call tokens (<start_function_call>call:...<end_function_call>).
- **Optimization:**
- Trained on a NVIDIA Tesla T4 GPU (16 GB VRAM).
- Effective batch size of 32 (micro-batch size 1 with 32 gradient accumulation steps).
- AdamW optimizer with FP16 mixed precision.
- Learning rate of 2e-5 with 36-step linear warmup and linear decay across 1 full epoch (1,214 steps).
- Zero-copy chunked cross-entropy (chunk_size=256) kept peak VRAM usage under 5.0 GB.
---

## Evaluation Benchmark

| Register / Language | Samples | Fn Acc (%) | Arg Acc (%) | Exact Match (EM) |
| :--- | :---: | :---: | :---: | :---: |
| **English** | 71 | 69.0% | 67.6% | 67.6% |
| **Burmese** | 47 | 44.7% | 44.7% | 44.7% |
| **Code-Switch** | 73 | 52.1% | 52.1% | 52.1% |
| **Overall** | **191** | **56.5%** | **56.0%** | **56.0%** |

- **Fn Acc (Function Accuracy):** Percentage of samples where all predicted tool names match ground truth.
- **Arg Acc (Argument Accuracy):** Percentage of samples where tool names AND all arguments match ground truth.
- **Exact Match (EM):** Strict end-to-end match.

### Per-Tool Accuracy
*Evaluated at tool invocation level (minimum 3 samples):*

| Tool Name | Correct | Total | Accuracy (%) |
| :--- | :---: | :---: | :---: |
| `turn_off_flashlight` | 12 | 12 | 100.0% |
| `check_traffic` | 3 | 3 | 100.0% |
| `open_website` | 4 | 4 | 100.0% |
| `open_wifi_settings` | 8 | 9 | 88.9% |
| `create_contact` | 30 | 35 | 85.7% |
| `turn_on_flashlight` | 15 | 18 | 83.3% |
| `navigate_to` | 13 | 16 | 81.2% |
| `send_email` | 18 | 26 | 69.2% |
| `show_map` | 17 | 29 | 58.6% |
| `create_calendar_event` | 33 | 61 | 54.1% |
| `set_volume_level` | 2 | 4 | 50.0% |
| `play_music` | 12 | 26 | 46.2% |
| `control_media` | 1 | 6 | 16.7% |

---

## Future Roadmap & Planned Improvements

As this is an initial release, future iterations are planned:

- **Dataset Expansion:**  
  Expanding [thantzinphyo/Hey-Function-Burmese](https://huggingface.co/datasets/thantzinphyo/Hey-Function-Burmese) with more conversational Burmese dialogue, regional slang, complex compound commands, and multi-turn function calls.
- **Improving Argument Extraction:**  
  Training for additional epochs (3–5 epochs) to boost Argument Accuracy (Arg Acc), enabling precise extraction of Burmese dates, times, currency amounts, and entity names into JSON arguments.
- **On-Device Quantization:**  
  Exporting to GGUF and LiteRT formats (int4 / int8 quantization) for offline execution inside Android and Flutter mobile apps.

---

## License

This project is licensed under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0), in accordance with the base model's [Gemma Terms of Use](https://ai.google.dev/gemma/terms).

---

## Citation & Resources

- **Model Repository:** [thantzinphyo/Hey-Function-270M](https://huggingface.co/thantzinphyo/Hey-Function-270M)
- **Dataset:** [thantzinphyo/Hey-Function-Burmese](https://huggingface.co/datasets/thantzinphyo/Hey-Function-Burmese)
- **Codebase:** [GitHub](https://github.com/thantzinphyo-gg/Hey-Function-Burmese)

```bibtex
@misc{hey_270m_alpha_2026,
  title={Hey-Function-270M: A Lightweight On-Device Burmese & English Function-Calling Agent},
  author={Thant Zin Phyo},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/thantzinphyo/Hey-Function-270M}}
}
```