Build Colab training notebook (Unsloth + LoRA)

[William-Hill]

Summary

Create a single "Run All" Colab notebook that trains fine-tuned models using Unsloth + LoRA on A100 GPU. Replace the MLX-based training/finetune.py with an Unsloth-backed wrapper.

Depends On

• Merge distillation-pipeline and shap-explainability to main #96

Design Principles

• Single "Run All" execution — no babysitting, no manual cell-by-cell
• Parameterized config cell — only cell the user edits (school, model sizes, tokens, epochs)
• Checkpoint and resume — SKIP_DOMAIN_ADAPTATION=True resumes from Phase 2 after disconnects
• Chat template alignment — uses tokenizer.apply_chat_template() throughout (D4BL critical lesson)

Tasks

• Create notebooks/training/bishop_state_fine_tuning.ipynb
• Cell 1: Config (SCHOOL, MODEL_SIZES=["4b","9b"], HF_TOKEN, epoch counts, skip flags)
• Cell 2+: Autonomous pipeline:
  • GPU detection + validation
  • pip install unsloth, trl, peft
  • Clone repo, load config.yaml
  • Phase 1: Domain adaptation (per model size) — LoRA r=16, all modules, 1 epoch, lr 2e-4, bf16
  • Phase 2: Task adapters (narrator, summarizer, explainer per model size) — task-specific LoRA config
  • Phase 3: GGUF export (q4_k_m) + upload to Drive/HF Hub
  • Summary: comparison table 4B vs 9B, recommend winner
• Replace training/finetune.py (MLX) with Unsloth-backed training/finetune.py
• Delete training/export.py (MLX Ollama export) — GGUF export moves into notebook
• Error handling: try/except per phase, save partial state before re-raising
• Loss curves saved as PNG alongside GGUFs
• Progress via tqdm + print (no interactive widgets)

Training Hyperparameters (from D4BL)

Parameter	Phase 1 (Domain)	Phase 2 (Tasks)
LoRA rank	16	8-16
LoRA alpha	32	16-32
Learning rate	2e-4	1e-4
Effective batch	32	16
Epochs	1	4-7
Sequence length	4096	4096-8192
Precision	bf16	bf16

Acceptance Criteria

• Notebook runs end-to-end on Colab A100 via "Run All" without manual intervention
• Produces GGUF files for all 3 tasks × 2 model sizes
• Prints comparison metrics table at end