Skip to content
View HarleyCoops's full-sized avatar
😁
Cooking what is next
😁
Cooking what is next

Block or report HarleyCoops

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
HarleyCoops/README.md

Christian H. Cooper

Quant trader · visual reasoning · modified GRPO · languages from single historical books

GitHub Hugging Face X Weights & Biases

3D Contribution Chart

I train small models to do hard reasoning — and to speak from books that almost no one still reads aloud.


Now

Modified GRPO → languages
Dakota1890 → Cree1865 → AutoScientist loops. One public-domain volume becomes grammar rules, verifiable rewards, and a community-correctable model.

Visual reasoning
Math-To-Manim (2.4k★) — if a model can render a recursive Archimedean solid, it can probably think.

Place & archive
Fortress · MONOLITH · Alberta LiDAR · Hector’s hand · Warre & Vavasour journals — terrain and ink as first-class data.


Visual Reasoning

Math-To-Manim Pythagorean animation

Math-To-Manim — text & images → epic math/physics animations & study notes
Manim · visual eval · code-gen reasoning · KimiK2Manim

Mythos QFT vertex diagram

Recursive geometry as eval
The 62-face recursive rhombicosidodecahedron is the stress test: nested polyhedra, exact coordinates, render-or-fail. Most LLMs die on the math alone.

Black-Scholes volatility surface

Black-Scholes volatility surface
The smile exotic desks actually trade against — CFA intuition rendered as motion, not a static chart.

Mythos learns Math-To-Manim

Animation as a reasoning eval: the model must plan geometry, write Manim, and survive the render-repair loop.


Generalized Learning via Modified GRPO

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 — the template

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.

Cree1865 — the generalization test

Watkins 1865 Cree dictionary title page

Watkins 1865 — A Dictionary of the Cree Language

Contact sheet of Watkins 1865 pages

Hypothesis: one historical volume is enough to bootstrap a correctable low-resource model.

  • Base: Qwen3-30B-A3B + LoRA
  • ~19.5k dictionary entries → ~38k RL tasks
  • Deterministic Cree reward ledger (exact / containment / orthography / F1 / length)
  • Live run: W&B hda2wqhl
  • Artifacts: model · Space

AutoScientist loops

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


Place, Terrain & Handwriting

Archival intelligence for the Canadian Rockies — maps that fly, journals that speak, places that reveal their people.

Alberta LiDAR

LiDAR demo

Map-driven LiDAR / DEM visualizer — pick a region, pull OpenTopography elevation, render multi-layer 3D terrain.

Live demo · lidar2 · contour

Three.js · OpenTopography · flyable topo

Warre & Vavasour

WV archival

Archival transcription dossiers and flyable map films for the Warre & Vavasour journals.

Live site · wv-archival

archival dossiers · map films

MONOLITH + FORTRESS

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.

Hector’s handwriting

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.


Hugging Face

Hugging Face

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)

Tape

Category Date Headline
Market Jul 16, 2026 Dallas Fed President Logan calls for 'modestly' higher interest rates
Market Jul 16, 2026 Short sellers load up against SpaceX as stock drops below IPO price
Market Jul 16, 2026 Sen. Warren says Trump's CFPB overhaul has cost Americans $26.5 billion
Market Jul 15, 2026 Anthropic moves closer to mega-IPO as bankers line up investor meetings
Market Jul 15, 2026 Fed Chairman Kevin Warsh's testimony to Senate banking committee hits on economy, interest rates
Finance Jul 16, 2026 US chip and memory stocks slide in fresh bout of Wall Street tumult
Finance Jul 16, 2026 Chinese AI start-up Moonshot to launch model challenging Anthropic’s lead
Finance Jul 16, 2026 A wealth tax in America? Not if Silicon Valley’s billionaires have their way
Finance Jul 16, 2026 The perks of parenting by spreadsheet
Finance Jul 16, 2026 Conflicts of interest are back — and more blatant than ever

Karpathy reply    Google Scholar citation

𝕏 · Hugging Face · W&B · Math-To-Manim · Dakota1890 · Cree1865

Pinned Loading

  1. Math-To-Manim Math-To-Manim Public

    Create Epic Math and Physics Animations & Study Notes From Text and Images.

    Python 2.4k 258

  2. StoneyNakoda StoneyNakoda Public

    A locally trained model of Stoney Nakoda has been developed and released. You can access the working model here or train your own instance.

    Python 10

  3. OneShotAquaRAT OneShotAquaRAT Public

    One click away from a locally downloaded, fine-tuned model, hosted on hugging face, with inference built in. In two hours.

    Jupyter Notebook 24 4

  4. nanochat561 nanochat561 Public

    Forked from karpathy/nanochat

    The best ChatGPT that $250 can buy.

    Python 6 2

  5. Dakota1890 Dakota1890 Public

    Using GRPO and a modified compositional reward function to train an opensource model on the 1890 Dakota Dictionary

    HTML 14

  6. TinyRecursiveInference TinyRecursiveInference Public

    Forked from SamsungSAILMontreal/TinyRecursiveModels

    Python