{ "cells": [ { "cell_type": "markdown", "id": "0", "metadata": {}, "source": [ "# 🌾 Sahel-Agri Voice AI — Fine-tune on Farmer Feedback\n", "\n", "**Run after collecting â‰Ĩ10 corrections in the Space.** \n", "First run? Use `bootstrap_repos.ipynb` instead to train the v0 Waxal adapter.\n", "\n", "This notebook fine-tunes the existing LoRA adapter using:\n", "- **Waxal baseline** (up to 500 samples) — keeps the model grounded\n", "- **Farmer corrections** (3× upsampled) — targeted improvement from real field use\n", "\n", "**Before running:** Runtime → Change runtime type → **T4 GPU**" ] }, { "cell_type": "code", "execution_count": null, "id": "1", "metadata": {}, "outputs": [], "source": [ "# Cell 1 — GPU check\n", "import subprocess\n", "result = subprocess.run(['nvidia-smi'], capture_output=True, text=True)\n", "if result.returncode != 0:\n", " raise RuntimeError('No GPU! Runtime → Change runtime type → T4 GPU')\n", "print(result.stdout[:500])\n", "print('✅ GPU ready')" ] }, { "cell_type": "code", "execution_count": null, "id": "2", "metadata": {}, "outputs": [], "source": [ "# Cell 2 — Install dependencies (matching Space versions)\n", "!pip install -q \\\n", " torch==2.11.0 torchaudio==2.11.0 \\\n", " transformers==5.5.0 datasets==4.8.4 \\\n", " accelerate==1.13.0 evaluate==0.4.2 \\\n", " huggingface-hub==1.9.0 peft==0.18.1 \\\n", " librosa==0.10.2 soundfile==0.12.1 \\\n", " jiwer==3.0.4 pyyaml==6.0.2\n", "print('✅ Packages installed')" ] }, { "cell_type": "code", "execution_count": null, "id": "3", "metadata": {}, "outputs": [], "source": [ "# Cell 3 — HuggingFace login\n", "# Colab: 🔑 icon (left sidebar) → Add new secret → name=HF_TOKEN\n", "# Kaggle: Add Data → add as Kaggle secret named HF_TOKEN\n", "import os\n", "try:\n", " from google.colab import userdata # type: ignore\n", " HF_TOKEN = userdata.get('HF_TOKEN')\n", "except Exception:\n", " HF_TOKEN = os.environ.get('HF_TOKEN', '')\n", "\n", "if not HF_TOKEN:\n", " raise ValueError('HF_TOKEN not found — see instructions above.')\n", "\n", "from huggingface_hub import login\n", "login(token=HF_TOKEN, add_to_git_credential=False)\n", "\n", "SPACE_REPO_ID = 'ous-sow/sahel-agri-voice'\n", "FEEDBACK_REPO_ID = 'ous-sow/sahel-agri-feedback'\n", "ADAPTER_REPO_ID = 'ous-sow/sahel-agri-adapters'\n", "# Must match what the Space uses — whisper-small for cpu-basic, whisper-large-v3-turbo for GPU.\n", "WHISPER_MODEL_ID = 'openai/whisper-small'\n", "TRAIN_LANG = 'bam' # ← change to 'ful' for Fula\n", "\n", "print(f'✅ Logged in | training language: {TRAIN_LANG}')" ] }, { "cell_type": "code", "execution_count": null, "id": "4", "metadata": {}, "outputs": [], "source": [ "# Cell 4 — Download Space code and feedback corrections\n", "import json, shutil, sys\n", "from pathlib import Path\n", "from huggingface_hub import snapshot_download, hf_hub_download\n", "\n", "# Get Space code (contains src/, configs/)\n", "space_dir = Path(snapshot_download(\n", " repo_id=SPACE_REPO_ID, repo_type='space', token=HF_TOKEN\n", "))\n", "sys.path.insert(0, str(space_dir))\n", "print(f'Space code: {space_dir}')\n", "\n", "# Download feedback corrections.jsonl\n", "jsonl_path = hf_hub_download(\n", " repo_id=FEEDBACK_REPO_ID,\n", " filename='corrections.jsonl',\n", " repo_type='dataset',\n", " token=HF_TOKEN,\n", ")\n", "with open(jsonl_path, encoding='utf-8') as f:\n", " all_records = [json.loads(l) for l in f if l.strip()]\n", "\n", "corrections = [\n", " r for r in all_records\n", " if r.get('is_correction') and r['language'] == TRAIN_LANG\n", "]\n", "print(f'Total feedback records : {len(all_records)}')\n", "print(f'Corrections for {TRAIN_LANG} : {len(corrections)}')\n", "\n", "if len(corrections) < 5:\n", " print('⚠ī¸ Very few corrections — consider collecting more before training.')\n", " print(' Training will proceed with Waxal only (corrections will be skipped).')" ] }, { "cell_type": "code", "execution_count": null, "id": "5", "metadata": {}, "outputs": [], "source": [ "# Cell 5 — Download feedback audio files from HF Dataset repo\n", "fb_audio_dir = Path('/tmp/sahel_feedback_audio')\n", "fb_audio_dir.mkdir(exist_ok=True)\n", "\n", "skipped = 0\n", "for rec in corrections:\n", " local_path = fb_audio_dir / Path(rec['audio_file']).name\n", " if local_path.exists():\n", " continue\n", " try:\n", " dl = hf_hub_download(\n", " repo_id=FEEDBACK_REPO_ID,\n", " filename=rec['audio_file'],\n", " repo_type='dataset',\n", " token=HF_TOKEN,\n", " )\n", " shutil.copy(dl, local_path)\n", " except Exception as e:\n", " skipped += 1\n", " print(f' skip {rec[\"audio_file\"]}: {e}')\n", "\n", "# Point records at local paths\n", "for rec in corrections:\n", " local = fb_audio_dir / Path(rec['audio_file']).name\n", " if local.exists():\n", " rec['audio_file'] = str(local)\n", "\n", "available = [r for r in corrections if Path(r['audio_file']).exists()]\n", "print(f'Downloaded {len(available)} / {len(corrections)} audio files (skipped {skipped})')" ] }, { "cell_type": "code", "execution_count": null, "id": "6", "metadata": {}, "outputs": [], "source": [ "# Cell 6 — Fine-tune: Waxal baseline + farmer corrections\n", "#\n", "# WhisperLoRATrainer.setup() loads Waxal (streaming).\n", "# merge_extra_data() materialises Waxal (up to 500 samples),\n", "# appends corrections (3× upsampled), shuffles the combined dataset.\n", "# train() runs standard Seq2SeqTrainer on the merged dataset.\n", "\n", "import os\n", "os.environ['HF_TOKEN'] = HF_TOKEN\n", "\n", "from src.training.trainer import WhisperLoRATrainer\n", "\n", "lang_config_map = {'bam': 'lora_bambara.yaml', 'ful': 'lora_fula.yaml'}\n", "base_cfg = str(space_dir / 'configs' / 'base_config.yaml')\n", "lang_cfg = str(space_dir / 'configs' / lang_config_map[TRAIN_LANG])\n", "output_dir = f'/tmp/sahel_adapter_{TRAIN_LANG}'\n", "\n", "# Override output_dir so adapter saves to /tmp on Colab\n", "import yaml\n", "with open(lang_cfg) as f:\n", " lang_config = yaml.safe_load(f)\n", "lang_config['output_dir'] = output_dir\n", "tmp_lang_cfg = f'/tmp/lora_{TRAIN_LANG}_tmp.yaml'\n", "with open(tmp_lang_cfg, 'w') as f:\n", " yaml.dump(lang_config, f)\n", "\n", "trainer = WhisperLoRATrainer(\n", " base_config_path=base_cfg,\n", " language_config_path=tmp_lang_cfg,\n", ")\n", "trainer.setup()\n", "\n", "if available:\n", " print(f'Merging {len(available)} corrections (×3) with Waxal baseline (cap=500)...')\n", " trainer.merge_extra_data(available, repeat=3, waxal_cap=500)\n", "else:\n", " print('No corrections available — training on Waxal only.')\n", "\n", "trainer.train()\n", "print(f'✅ Training complete — adapter at {output_dir}')" ] }, { "cell_type": "code", "execution_count": null, "id": "7", "metadata": {}, "outputs": [], "source": [ "# Cell 7 — Push adapter to HF Model repo\n", "from huggingface_hub import HfApi\n", "api = HfApi(token=HF_TOKEN)\n", "\n", "path_in_repo = 'adapters/bambara' if TRAIN_LANG == 'bam' else 'adapters/fula'\n", "n_corrections = len(available)\n", "\n", "api.upload_folder(\n", " folder_path=output_dir,\n", " repo_id=ADAPTER_REPO_ID,\n", " repo_type='model',\n", " path_in_repo=path_in_repo,\n", " commit_message=(\n", " f'Fine-tune {TRAIN_LANG}: Waxal baseline + {n_corrections} farmer corrections'\n", " ),\n", ")\n", "print(f'✅ Pushed to {ADAPTER_REPO_ID}/{path_in_repo}')\n", "print('\\nNext: Space → Tab 3 → Reload Adapters from Hub')" ] }, { "cell_type": "code", "execution_count": null, "id": "8", "metadata": {}, "outputs": [], "source": [ "# Cell 8 — Sanity check: compare WER before vs after adapter\n", "import random, torch, librosa, jiwer\n", "from transformers import WhisperForConditionalGeneration, WhisperProcessor\n", "from peft import PeftModel\n", "\n", "if not available:\n", " print('No test samples — skipping sanity check.')\n", "else:\n", " test_rec = random.choice(available)\n", " print(f'Audio : {Path(test_rec[\"audio_file\"]).name}')\n", " print(f'Expected : {test_rec[\"corrected_text\"]}')\n", " print(f'Pre-train: {test_rec[\"whisper_output\"]}')\n", "\n", " # Load base + adapter\n", " processor = WhisperProcessor.from_pretrained(WHISPER_MODEL_ID, token=HF_TOKEN)\n", " base = WhisperForConditionalGeneration.from_pretrained(\n", " WHISPER_MODEL_ID, torch_dtype=torch.float16, token=HF_TOKEN\n", " ).to('cuda')\n", " model = PeftModel.from_pretrained(base, output_dir).eval()\n", "\n", " audio_np, _ = librosa.load(test_rec['audio_file'], sr=16000, mono=True)\n", " feats = processor.feature_extractor(\n", " audio_np, sampling_rate=16000, return_tensors='pt'\n", " ).input_features.half().to('cuda')\n", "\n", " with torch.no_grad():\n", " ids = model.generate(feats, max_new_tokens=256)\n", " result = processor.batch_decode(ids, skip_special_tokens=True)[0].strip()\n", " print(f'Post-train: {result}')\n", "\n", " ref = test_rec['corrected_text']\n", " wer_before = jiwer.wer(ref, test_rec['whisper_output']) if test_rec.get('whisper_output') else 1.0\n", " wer_after = jiwer.wer(ref, result)\n", " print(f'\\nWER before: {wer_before:.1%} → WER after: {wer_after:.1%}')\n", " if wer_after < wer_before:\n", " print('✅ Adapter improved transcription quality!')\n", " else:\n", " print('ℹī¸ No improvement on this single sample — collect more corrections and retrain.')" ] } ], "metadata": { "accelerator": "GPU", "colab": { "gpuType": "T4", "provenance": [] }, "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "name": "python", "version": "3.10.0" } }, "nbformat": 4, "nbformat_minor": 5 }