Update README.md

demo v3: deterministic textbook diagrams — neural networks + gradient descent, clean templated boards

Improve whiteboard density and callouts

Tighten first-load UI layout

Keep Fable directive notes within three steps

Fix current-source lessons and compact board layout

Enrich science whiteboard templates

Enrich whiteboard diagrams and responsive board

Clarify orbit and gradient diagrams

Improve whiteboard diagram quality

demo v2.2: grouped text stays together, physically accurate orbit trajectory

text groups merge into notes blocks; curves must be shape-true in screen coords

demo v2.1: expanded board fits the frame, caption strip always visible

expanded board fits the viewport (caption strip stays visible)

sync: env-gated tts endpoint (inert without TUTORI_TTS_DEMO)

demo video v2: teacher's-studio UI, MiniCPM coach, live research, expanding board, 16.9B stack

profile personalizes, never vetoes: free topic switches, honest coach goals, notes clear on new lesson

no more slivers: min shape dims + vertical label fit

badges flow left, clear of the HF mini-header pill

goal/start dots snap to the curve's actual lowest/highest point

polygon labels fit the shape interior (chord-width shrink)

board polish: highlights hug measured text, arrows chase displaced labels, hypotenuse label on longest side

above-the-fold board: header:mini + compact chalk header; field notes article linked

fixes: shrink-to-fit shape labels, dark-mode-proof text, responsive stack, coach echo guard

v2: teacher's-studio design, MiniCPM5 planner+coach, expanding board, field notes

README: full Well-Tuned scorecard — two artists, live A/B, honest retirement

diagnostic: defer-phase timing logs

gemma artist: fit 59s budget — TTS overlaps stream, 3-4 steps, shorter renders

dual board artist: TUTORI_BOARD_MODEL=gemma mounts LoRA on the teacher (ZeroGPU-compatible)

nemotron: stop on <|im_end|> + first-balanced-JSON parse (model output verified good)

nemotron: enable_thinking=False to match tuned adapter training rendering

README: honest planner-fallback note (ZeroGPU Mamba prefill constraint)

README: Well-Tuned badge — published board-artist LoRA + dataset, honest flag-off story

revert to torch 2.11: mamba kernels can't JIT within ZeroGPU turn budget; board model stays flag-off

pin torchaudio/torchvision to torch 2.10 ABI

torch 2.10 + fused mamba kernels: naive Mamba scan OOMs ZeroGPU on Nemotron generate

fix nemotron_generate: unwrap BatchEncoding from apply_chat_template (also un-breaks Nemotron planner)

debug: full traceback on board render failure

pin peft 0.18.1 (0.19.x adapter-load path needs newer transformers WeightConverter API)

board model: mount LoRA adapter lazily inside GPU worker (ZeroGPU has no CUDA at import)

board-model experiment: flagged Nemotron LoRA board renderer (off by default)

Upload board_check.py with huggingface_hub

Upload engine.py with huggingface_hub

Upload folder using huggingface_hub

Upload README.md with huggingface_hub

Upload README.md with huggingface_hub

Upload tutori_demo.mp4 with huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Delete pre-requirements.txt with huggingface_hub