Pipeline
idle
1
Fetch Training Data
Pull corpora and model cards from HuggingFace + GitHub (2026 sources).
2
Train Model
GRPO policy optimization + VAE filter on a bigram model.
3
Export Bundle
Download a HuggingFace-ready .zip with model, tests, diagrams.
Metrics
Steps
0
Loss
—
Reward
—
Vocab
0
States
0
Cache Hit
0%
Settings
Group Size8
KL Penalty0.100
VAE Latent Dim (metadata)32
Filter Threshold (metadata)0.70
Reference Architecture
Reference only — metadata used in export. Actual model is character-level bigram. GRPO training uses group size and KL penalty; VAE fields are recorded in export metadata only.
[api-fanout] Idle. Click "Fetch" in the pipeline to begin.
[api-fanout] Endpoints (2026, HuggingFace + GitHub only):
raw.githubusercontent.com/karpathy/makemore/master/names.txt corpus
huggingface.co/api/models?search=smollm2&limit=10 SmolLM2 models
huggingface.co/api/models?search=qwen2.5-0.5b&limit=5 Qwen2.5 models
huggingface.co/HuggingFaceTB/SmolLM2-135M/raw/main/README.md SmolLM2 card
huggingface.co/allenai/OLMo-2-0425-1B/raw/main/README.md OLMo-2 card
api.github.com/repos/karpathy/makemore Karpathy repo
Progress
idle
4x Harmonic Oscillator
superposition
Four incommensurate frequencies (1, phi, 2, pi). Coherent equal-weight superposition — |c_i|² is invariant under Hamiltonian evolution, so probabilities stay at 0.25 each. Measurement collapse drives the export seed.