Upload minimum_training_script.ipynb

Delete notebooks/train_colab.ipynb

Upload notebooks/train_colab.ipynb with huggingface_hub

Upload architecture.png

Delete assets/architecture.png

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload assets/architecture.png with huggingface_hub

Add training results visualization with reward trend chart

Increase training scale: more steps, episodes, and SFT epochs

Pre-format SFT dataset as text column, drop formatting_func

Fix pickling error in SFT formatting_func closure

Fix SFT formatting_func to return list of strings

Fix SFT: set completion_only_loss=False for formatting_func compat

Fix SFT warm start: add formatting_func for Unsloth SFTTrainer

Cap prompt generation at 512 tokens and add version print

Add SFT warm start before GRPO and DB connectivity init check

Merge pull request #13 from KarlLearnsAI/main

Merge pull request #12 from KarlLearnsAI/claude/ai-oversight-system-ThVHS

Merge pull request #11 from KarlLearnsAI/main

Move supabase to core dependencies

Add train.sh startup script and assets folder

Fix Gradio launch to bind 0.0.0.0 for HF Spaces

Replace app with static architecture overview (no LLM calls on startup)

Add HF Spaces config metadata to README

Merge pull request #10 from KarlLearnsAI/claude/ai-oversight-system-ThVHS

Switch Llama 3.1 8B to ungated unsloth mirror

Merge pull request #9 from KarlLearnsAI/claude/ai-oversight-system-ThVHS

Add local model inference backend for Layer 2

Increase max completion length from 512 to 2048

Add 502/504 and hyphenated Time-out to retry list

Merge pull request #8 from KarlLearnsAI/claude/ai-oversight-system-ThVHS

Add retry with exponential backoff for HF Inference API calls

Merge pull request #7 from KarlLearnsAI/claude/ai-oversight-system-ThVHS

Make Supabase uploads incremental — upload after every step

Add supabase to Dockerfile pip install

Merge pull request #6 from KarlLearnsAI/claude/ai-oversight-system-ThVHS

Add Supabase upload for training results (Storage + DB)

Add raw training summary output and adjust training scale

Improve reward function to break refuse-everything local minimum and scale training

Merge pull request #5 from KarlLearnsAI/claude/ai-oversight-system-ThVHS

Add volume verification, fsync, and stdout fallback for training outputs

Merge pull request #4 from KarlLearnsAI/claude/ai-oversight-system-ThVHS

Update output paths to use persistent volume at /workspace/output

Clean up dead code, unused imports, and move hardcoded values to config.yaml

Add --llm-agent and other legacy CLI flags for backwards compatibility

Merge pull request #3 from KarlLearnsAI/claude/ai-oversight-system-ThVHS

Reduce episodes_per_candidate from 5 to 3

Reduce GRPO training params to minimum: 2 candidates, 5 steps, 5 episodes

Centralize all training params in config.yaml (single source of truth)