chakibmed/whispbook-functiongemma-270m-speaker-attribution-mlx-lora-v4-r010-b6-step850

Whispbook FunctionGemma Speaker Attribution MLX LoRA v4 Step 850

This repository contains the selected MLX-LM LoRA adapter for the Whispbook book-dialogue speaker attribution experiment.

• —Base model: mlx-community/functiongemma-270m-it-4bit
• —Task: FunctionGemma tool-call speaker attribution
• —Selected checkpoint: 0000850_adapters.safetensors
• —Local run: functiongemma_not_in_candidates_v4_ratio010_b6
• —Training data recipe: not_in_candidates_ratio=0.10, batch_size=6, learning_rate=5e-6, 900 planned steps

The uploaded adapters.safetensors file is copied from checkpoint 0000850_adapters.safetensors, not from the final step 900 adapter.

Evaluation

Full 500-row generation evaluation for the selected checkpoint:

Interpretation: this checkpoint has very high recall for not_in_candidates, but it can overuse the fallback option on known-speaker rows. It is best treated as the high-recall fallback experiment, not as a finished production speaker attribution model.

See:

• —eval/eval_0000850_full500_summary.json
• —report/README_metrics.md
• —report/training_report.html
• —report/checkpoint_metrics.json

Loading With MLX-LM

uv run --with mlx-lm python - <<'PY'
from mlx_lm import load, generate

model, tokenizer = load(
    "mlx-community/functiongemma-270m-it-4bit",
    adapter_path="chakibmed/whispbook-functiongemma-270m-speaker-attribution-mlx-lora-v4-r010-b6-step850",
)

prompt = "<bos><start_of_turn>developer\n...\n<start_of_turn>model\n"
print(generate(model, tokenizer, prompt, max_tokens=96))
PY

Intended Use

This adapter is for research on public-domain-derived book dialogue speaker attribution. It is not for audio generation, voice assignment, or storing raw copyrighted books.