Tapasya needs a desktop browser

WebGPU-accelerated training works on desktop Chrome, Edge, or Safari. Open this URL on your laptop.

Tapasya
Train a language model in your browser — three stages, no server, no account, no data leaving your device.

Guided tour

Walk through all three stages with teaching copy explaining each step.

Free mode

All three stages open from the start. Change any setting.

b2 — LoRA fine-tuning

Fine-tune SmolLM-135M on your documents with rank-8 LoRA. Load files, train, export adapters.

Tapasya runs three progressively sophisticated training regimes on your text corpus — byte-level minimal, byte-level with more compute, and BPE-tokenized — using only forward passes (no backprop). Each stage's output motivates the next stage's addition.

The algorithm is EGGROLL — evolution strategies without backprop. Your corpus never leaves the tab.

Tapasya / b2 idle
Model
SmolLM-135M
Not loaded
Documents
Drop .txt / .md files or click to browse
Settings
LoRA targets
q_proj k_proj v_proj o_proj down_proj gate_proj up_proj
Loss
Generated samples (last 10)
Samples appear after each training batch.