tinylmtune 0.0.3

A lightweight Python library that automates TinyBERT fine-tuning with Genetic Algorithm hyperparameter optimisation

tinyLMTune

Genetic-Algorithm-Optimised TinyBERT Fine-Tuning in one function call.

tinyLMTune automates the full pipeline: (1) dataset building
(2) GA hyperparameter search (3) TinyBERT fine-tuning (4) model export

Bring your own data or let it generate synthetic training data automatically.

Installation

pip install tinylmtune

Quick Start

Option 1 — Bring your own data (recommended)

from tinylmtune import optimize_slm, TinyInference

my_data = [
    {"text": "Loved this product!", "label": "positive"},
    {"text": "Broke after one day.", "label": "negative"},
    {"text": "Works fine, nothing special.", "label": "neutral"},
    # ... at least 50+ records recommended
]

best = optimize_slm(
    task="classification",
    user_data=my_data,
    output_dir="my_model",
)

model = TinyInference("my_model")
print(model.predict("Absolutely amazing quality!"))

Option 2 — Use a JSONL file

best = optimize_slm(
    task="classification",
    corpus_path="my_data.jsonl",   # one JSON object per line
    output_dir="my_model",
)

Option 3 — Auto-generate synthetic data (default fallback)

If neither user_data nor corpus_path is provided, tinyLMTune generates synthetic training data via a local Ollama/Mistral model:

best = optimize_slm(
    task="classification",
    corpus_prompt="Generate movie-review sentiment examples",
    n_examples=200,
    output_dir="my_model",
)

Data Format

Each record is a dict (or a JSON line in a .jsonl file). The required keys depend on the task:

Task	Required keys	Example
classification	text, label	{"text": "Great film!", "label": "positive"}
summarization	text, summary	{"text": "Long article...", "summary": "Short version."}
qna	question, answer	{"question": "What is X?", "answer": "X is..."}
generation	prompt, completion	{"prompt": "Once upon a", "completion": "time there was..."}
ner	text, entities	{"text": "John in NYC", "entities": [{"text": "John", "label": "PER", "start": 0, "end": 4}]}

Records missing required keys are skipped with a warning. If all records are invalid, an error is raised showing the expected format.

Public API

Only two symbols are exported:

Symbol	Purpose
optimize_slm()	Full train pipeline — returns best config
TinyInference	Load & predict with a saved model

All internal modules live in tinylmtune._internal and are not part of the public API.

Project Structure

tinylmtune/
├── __init__.py              # Public API: optimize_slm, TinyInference
├── _internal/               # Private — do not import directly
│   ├── __init__.py
│   ├── constants.py         # Model names, task mappings, GA search space
│   ├── cleaner.py           # Text cleaning utilities
│   ├── corpus_gen.py        # Synthetic data generation (Ollama/Mistral)
│   ├── dataset.py           # list[dict] / JSONL → HuggingFace Dataset
│   ├── model_builder.py     # TinyBERT model instantiation
│   ├── trainer.py           # HuggingFace Trainer wrapper
│   ├── ga_optimizer.py      # Genetic algorithm hyperparameter search
│   ├── inference.py         # TinyInference class + save_best_model
│   └── pipeline.py          # optimize_slm() orchestrator
├── setup.py
└── README.md