🧬 EvoTransformer — Sovereign Kreol LLM (live)
Chat with a language model for Kreol Morisien that was evolved, not hand-designed, and trained from scratch — then watch the engine build architectures in 🧬 Live evolution, and see how it evolves itself from gated feedback in 🔁 Self-evolution. Every number is measured by training real models; nothing is simulated.
💬 You're chatting with YOUR sovereign model —
HemanM/evotransformer-kreol, the EvoTransformer architecture you evolved and trained from scratch on Kreol Morisien. No API, no borrowed base — every weight is yours.Ask it a question in Kreol (tap an example below). It's a small model, so replies are short and sometimes imperfect — but it should answer, not ramble. The first message loads the model (~20s).
Tap an example below, or type a Kreol question and press Enter.
Try one (one click → it answers):
Click Run to train a population of architectures live and select the best by measured fitness.
Baseline vs evolved winner (measured)
Model | Params | tok/s | val_loss | fitness |
|---|---|---|---|---|
Best fitness by generation
generation | best_fitness |
|---|---|
🔁 How it evolves itself — safely
The headline is "a transformer that evolves itself from feedback." For that to be defensible (not a Tay-style liability), the loop has two guarantees:
1. Garbage can't get in. Every input passes a quality gate that scores it on transparent, training-free signals — length, letter ratio, character entropy, consonant runs (keysmash), vowels-per-word, and token fertility against the model's own tokenizer (does it even look like the languages it knows?). Each rejection states its reason. (Try the 👍/👎 on the Chat tab — it shows the gate's verdict on your input, live.)
2. The model can't get worse. A re-evolved challenger replaces the champion only if it measurably beats it on held-out perplexity. A round improves the model or changes nothing — it can never degrade it.
So learning is batched and gated, not live token-by-token: the deployed model is frozen between promotions, by design. Run the whole loop yourself (CPU, ~1–2 min):
python scripts/self_evolve_demo.py
Full writeup: docs/SELF_EVOLUTION.md. In the lineage of Google's Evolved
Transformer (2019), specialised to run GPU-poor and to close the feedback loop.
The USP — a sovereign AI you own end-to-end
As far as can be found, the first sovereign, instruction-following assistant for Kreol Morisien — prior public Kreol work is machine translation, not an owned generative model. Built by an architecture-evolution engine you own end-to-end. Three moats:
- Measured architecture evolution — the engine trains and measures many candidate architectures and selects the best; it never hand-designs or simulates. In the lineage of Google's Evolved Transformer (2019), specialised to run GPU-poor and to close a gated feedback loop.
- Language + sovereignty — Kreol Morisien (~1.3M speakers) is unserved by frontier models; this is a replicable template for any underserved language or sovereignty-seeking nation.
- Full-stack ownership — own the tokenizer, corpus, weights, evals and runtime. No API, no per-token cost, on-prem data residency.
Honest scope
A small (~40M–350M-class) sovereign model — useful for translation, drafting and grounded Q&A in en/fr/Kreol, not a frontier reasoning model. Evolution buys efficiency-per-FLOP, not scaling laws; the win is an owned, efficient, sovereign model with an audit trail. The Kreol is still being refined with native-speaker review.
Source & plan: github.com/drhemanm/EvoTransformerV11
— docs/INVESTOR_BRIEF.md, docs/SELF_EVOLUTION.md. 78 tests + 22 smoke checks pass on CPU.