README.md

---
title: Christian H. Cooper
emoji: ⬛
colorFrom: green
colorTo: indigo
sdk: static
pinned: false
license: mit
---

<div align="center">

# Christian H. Cooper
### ML engineer · modified GRPO · visual reasoning · languages from single books

<p>
  <a href="https://github.com/HarleyCoops"><img src="https://img.shields.io/badge/GitHub-HarleyCoops-181717?style=flat-square&logo=github" alt="GitHub"></a>
  <a href="https://x.com/christiancooper"><img src="https://img.shields.io/badge/X-@christiancooper-000000?style=flat-square" alt="X"></a>
  <a href="https://wandb.ai/christian-cooper-us"><img src="https://img.shields.io/badge/W%26B-experiments-FFBE00?style=flat-square&logo=weightsandbiases&logoColor=black" alt="W&B"></a>
  <a href="https://github.com/HarleyCoops/Math-To-Manim"><img src="https://img.shields.io/badge/Math--To--Manim-2.4k%20stars-d97757?style=flat-square" alt="Math-To-Manim"></a>
</p>

<sub>Quant trader turned ML engineer. I train small (and not-so-small) models with <b>deterministic compositional rewards</b> — then publish the adapters, datasets, Spaces, and W&B ledgers.</sub>

</div>

---

## Run Stats (live research ledger)

| Run | Model | Steps | Batch / Group | Headline metric | W&B |
|---|---|---|---|---|---|
| **Cree1865** `hda2wqhl` | Qwen3-30B-A3B + LoRA | **800** | 16 / 8 | composite **0.15 → ~0.30**; length channel → **~0.95** | [run](https://wandb.ai/christian-cooper-us/thinking-machines-qwen3-30b/runs/hda2wqhl) |
| **Dakota 35B** `owf98569` | Qwen3.6-35B-A3B GRPO | **199** | 48 / 16 | composite **0.166 → 0.230**; char-overlap **0.14 → 0.40**; affix → **1.00** | [run](https://wandb.ai/christian-cooper-us/dakota-rl-grammar/runs/owf98569) |
| **Dakota 0.6B-400** | Qwen3-0.6B GRPO | **400** | — | **+150%** reward; **90%** of gains in first **21%** of steps | [model](https://huggingface.co/HarleyCooper/Qwen3-0.6B-Dakota-Grammar-RL-400) |

<div align="center">
<img src="https://huggingface.co/HarleyCooper/Cree1865/resolve/main/assets/cree_reward_progression.png" alt="Cree composite reward progression" width="48%"/>
<img src="https://huggingface.co/HarleyCooper/Qwen3.6-35B-A3B-Dakota1890-GRPO/resolve/main/assets/qwen36_reward_progression.png" alt="Dakota 35B reward progression" width="48%"/>
<br/>
<sub>Left: Cree1865 composite reward · Right: Dakota 35B GRPO composite reward</sub>
</div>

<div align="center">
<img src="https://huggingface.co/HarleyCooper/Cree1865/resolve/main/assets/cree_reward_components.png" alt="Cree reward channels" width="48%"/>
<img src="https://huggingface.co/HarleyCooper/Qwen3.6-35B-A3B-Dakota1890-GRPO/resolve/main/assets/qwen36_reward_components.png" alt="Dakota reward channels" width="48%"/>
<br/>
<sub>Per-channel ledgers — no LLM judge. Every term is code-checkable.</sub>
</div>

---

## Thesis

```text
one public-domain volume
  → VLM extraction (preserve orthography)
  → executable grammar / dictionary rules
  → thousands of verifiable RL tasks
  → modified GRPO (deterministic reward ledger)
  → published adapter + Space
  → community correction (the real second stage)
```

**Generalized learning via modified GRPO:** qualitative linguistic goals become measurable primitives (orthography, affixes, containment, F1, length). The same loop that proved on Dakota1890 now runs on Cree1865 — and AutoScientist cycles try to improve the data, not just the weights.

---

## Featured Models

| Artifact | What it is |
|---|---|
| [`Cree1865`](https://huggingface.co/HarleyCooper/Cree1865) | Watkins 1865 Cree dictionary adapter · GRPO · Tinker |
| [`Qwen3.6-35B-A3B-Dakota1890-GRPO`](https://huggingface.co/HarleyCooper/Qwen3.6-35B-A3B-Dakota1890-GRPO) | Latest Dakota GRPO endpoint (35B) |
| [`Qwen3-30B-Dakota1890`](https://huggingface.co/HarleyCooper/Qwen3-30B-Dakota1890) | 30B MoE Dakota LoRA |
| [`Qwen3-0.6B-Dakota-Grammar-RL-400`](https://huggingface.co/HarleyCooper/Qwen3-0.6B-Dakota-Grammar-RL-400) | Tiny-model grammar proof · +150% reward |
| [`nanochat-AquaRat`](https://huggingface.co/HarleyCooper/nanochat-AquaRat) | Algebraic reasoning RL |
| [`Qwen.5B-OpenR1Math`](https://huggingface.co/HarleyCooper/Qwen.5B-OpenR1Math) | Open-R1 style math on 0.5B |
| [`nanochat561`](https://huggingface.co/HarleyCooper/nanochat561) | Full-stack nanochat training card |

---

## Visual Reasoning

<div align="center">
<a href="https://github.com/HarleyCoops/Math-To-Manim">
  <img src="https://raw.githubusercontent.com/HarleyCoops/Math-To-Manim/main/docs/assets/r1-pythagorean-tweet.gif" alt="Math-To-Manim Pythagorean animation" width="85%"/>
</a>
<br/>
<b><a href="https://github.com/HarleyCoops/Math-To-Manim">Math-To-Manim</a></b> (2.4k★) — animation as a reasoning eval<br/>
<sub>If a model can plan geometry, write Manim, and survive the render-repair loop, it can probably think.</sub>
</div>

<div align="center">
<img src="https://raw.githubusercontent.com/HarleyCoops/Math-To-Manim/main/docs/assets/mythos-learns-math-to-manim.png" alt="Mythos learns Math-To-Manim" width="70%"/>
</div>

---

## Languages from Single Books

<table>
<tr>
<td width="42%" align="center" valign="top">

<a href="https://huggingface.co/HarleyCooper/Cree1865">
  <img src="https://raw.githubusercontent.com/HarleyCoops/Cree1865/main/docs/story/title_page.png" alt="Watkins 1865 title page" width="100%"/>
</a>
<sub>Watkins 1865 · Cree dictionary</sub>

</td>
<td width="58%" valign="top">

### Cree1865
Hypothesis test: **one historical volume** is enough to bootstrap a correctable low-resource model.

- ~19.5k entries → ~38k RL tasks  
- Deterministic Cree reward ledger  
- Live Space: [`Cree1865-Tinker-Inference`](https://huggingface.co/spaces/HarleyCooper/Cree1865-Tinker-Inference)  
- Not a fluency claim — a community-correctable endpoint

### Dakota1890 ladder
0.6B proof → 30B LoRA → **35B GRPO**.  
Source: Riggs 1890 grammar & dictionary.  
Demo: [`Dakota-.6B`](https://huggingface.co/spaces/HarleyCooper/Dakota-.6B)

</td>
</tr>
</table>

---

## Datasets & Spaces

**Datasets:** [`dakota-bilingual-qa`](https://huggingface.co/datasets/HarleyCooper/dakota-bilingual-qa) · [`adaption-dakota-english-qa`](https://huggingface.co/datasets/HarleyCooper/adaption-dakota-english-qa) · [`synthetic_stoney_data`](https://huggingface.co/datasets/HarleyCooper/synthetic_stoney_data) (68.8k) · [`StoneyNakoda`](https://huggingface.co/datasets/HarleyCooper/StoneyNakoda) · [`StoneyNakoda45k`](https://huggingface.co/datasets/HarleyCooper/StoneyNakoda45k)

**Spaces:** [`Cree1865-Tinker-Inference`](https://huggingface.co/spaces/HarleyCooper/Cree1865-Tinker-Inference) · [`Dakota-.6B`](https://huggingface.co/spaces/HarleyCooper/Dakota-.6B) · [`StoneyApp`](https://huggingface.co/spaces/HarleyCooper/StoneyApp) · [`AskAboutMath2Manim`](https://huggingface.co/spaces/HarleyCooper/AskAboutMath2Manim)

---

## Quickstart

```python
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)
```

Treat outputs as **first attempts** for community correction — not authorities.

---

<div align="center">
<sub>
  <a href="https://github.com/HarleyCoops">GitHub</a> ·
  <a href="https://wandb.ai/christian-cooper-us">W&B</a> ·
  <a href="https://github.com/HarleyCoops/Dakota1890">Dakota1890</a> ·
  <a href="https://github.com/HarleyCoops/Cree1865">Cree1865</a> ·
  <a href="https://github.com/HarleyCoops/Math-To-Manim">Math-To-Manim</a>
</sub>
</div>