Fine‑Tune Trainer (Hugging Face)

Overview

• Fine‑tunes distilbert-base-uncased on dair-ai/emotion using HF Trainer.
• Provides a tiny CLI for inference; optional Weights & Biases logging.

Structure

• src/train.py: Fine‑tune + save best to runs/emotion-distilbert/best.
• src/infer.py: CLI inference (default uses Hub model ID).
• notebooks/main.ipynb: Exploration/experiments.
• pyproject.toml: Project metadata and deps.
• .env: Optional env vars (e.g., WANDB_API_KEY).
• runs/: Training outputs; wandb/: local W&B logs.

Setup

• Python >=3.12, uv installed (https://docs.astral.sh/uv/)
• Install deps: uv sync (creates/uses .venv and lockfile)
• Run without activating venv: prefix commands with uv run
• Optional logging: uv pip install wandb then wandb login

Train

• Command: uv run python src/train.py
• Saves model + tokenizer to runs/emotion-distilbert/best and logs metrics.

Infer

• From Hub: uv run python src/infer.py "I feel great today!"
• Use local checkpoint: set MODEL_ID = "runs/emotion-distilbert/best" in src/infer.py.