Quant trader · visual reasoning · modified GRPO · languages from single historical books
I train small models to do hard reasoning — and to speak from books that almost no one still reads aloud.
|
Modified GRPO → languages |
Visual reasoning |
Place & archive |
Math-To-Manim — text & images → epic math/physics animations & study notes
Manim · visual eval · code-gen reasoning · KimiK2Manim
Recursive geometry as eval |
Black-Scholes volatility surface |
Animation as a reasoning eval: the model must plan geometry, write Manim, and survive the render-repair loop.
The bet: deterministic, compositional rewards beat vague LLM-as-judge loops for qualitative tasks — especially when the “ground truth” is a single historical book.
one public-domain volume
→ VLM extraction (preserve orthography exactly)
→ executable grammar / dictionary rules
→ thousands of verifiable RL tasks
→ modified GRPO (no LLM judge)
→ published adapter
→ community correction (the real second stage)
Dakota1890 · Riggs 1890 grammar & dictionary · reward = orthography + affixes + semantics
| Model | Scale | Method | Link |
|---|---|---|---|
| Qwen3-0.6B-Dakota-Grammar-RL | 0.6B | GRPO proof | HF |
| Qwen3-0.6B-Dakota-Grammar-RL-400 | 0.6B | +150% reward | HF |
| Qwen3-30B-Dakota1890 | 30B MoE | Tinker LoRA | HF |
| Qwen3.6-35B-A3B-Dakota1890-GRPO | 35B | latest GRPO | HF |
The model is not always right. That is not the point. It has read the 1890 book cover to cover. Descendants of Dakota speakers teach it the rest — the StoneyNakoda community-in-the-loop stage.
Watkins 1865 — A Dictionary of the Cree Language |
Hypothesis: one historical volume is enough to bootstrap a correctable low-resource model.
|
Recursive data ↔ model cycles on the Dakota seed (private contest workspace).
Seed → Adaption upload → cycle manifest → deterministic QA harness → public release artifacts — show the loop, not only a chatbot.
Also in the GRPO family: StoneyNakoda · smolThinker-.5B · OneShotAquaRAT · Qwen3-RailroadEngineer1959-RL
Archival intelligence for the Canadian Rockies — maps that fly, journals that speak, places that reveal their people.
|
Map-driven LiDAR / DEM visualizer — pick a region, pull OpenTopography elevation, render multi-layer 3D terrain.
|
Archival transcription dossiers and flyable map films for the Warre & Vavasour journals.
|
|
Private research line: MONOLITH is a real-DEM terrain engine (Emerald Lake default, 1902/1921 historic sheet overlays, align-by-click georef). FORTRESS is the place-first Whyte Museum pipeline for Fortress Mountain / Kananaskis — location → archives → artifacts → people → story clusters. Repos are private today — no public Pages deploy yet. |
Private HTR line on Sir James Hector’s 1858 Kicking Horse field journals (Palliser / MS-0443): VLM draft → human gold → TrOCR / Qwen3-VL LoRA → CER/WER. Public companion is the Warre & Vavasour site above. Hector1858 + GenericOCR remain private while rights/gold-set work continues. |
| Artifact | What it is |
|---|---|
Qwen3.6-35B-A3B-Dakota1890-GRPO |
Latest Dakota GRPO endpoint |
Cree1865 |
Watkins 1865 Cree adapter |
Qwen3-0.6B-Dakota-Grammar-RL |
Tiny-model grammar proof |
nanochat-AquaRat |
Algebraic reasoning RL |
Qwen.5B-OpenR1Math |
Open-R1 style math, 0.5B |
dakota-bilingual-qa |
Dakota–English QA seed |
synthetic_stoney_data |
68.8k Stoney rows |
Cree1865-Tinker-Inference |
Live Cree sampler Space |
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = "Qwen/Qwen3-30B-A3B-Instruct-2507"
adapter = "HarleyCooper/Cree1865"
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", torch_dtype="auto")
tok = AutoTokenizer.from_pretrained(base)
model = PeftModel.from_pretrained(model, adapter)
𝕏 · Hugging Face · W&B · Math-To-Manim · Dakota1890 · Cree1865