README.md

metadata

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

Christian H. Cooper

ML engineer · modified GRPO · visual reasoning · languages from single books

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

---

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
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
Dakota 0.6B-400	Qwen3-0.6B GRPO	400	—	+150% reward; 90% of gains in first 21% of steps	model

_{Left: Cree1865 composite reward · Right: Dakota 35B GRPO composite reward}

_{Per-channel ledgers — no LLM judge. Every term is code-checkable.}

---

Thesis

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	Watkins 1865 Cree dictionary adapter · GRPO · Tinker
Qwen3.6-35B-A3B-Dakota1890-GRPO	Latest Dakota GRPO endpoint (35B)
Qwen3-30B-Dakota1890	30B MoE Dakota LoRA
Qwen3-0.6B-Dakota-Grammar-RL-400	Tiny-model grammar proof · +150% reward
nanochat-AquaRat	Algebraic reasoning RL
Qwen.5B-OpenR1Math	Open-R1 style math on 0.5B
nanochat561	Full-stack nanochat training card

---

Visual Reasoning

Math-To-Manim (2.4k★) — animation as a reasoning eval
_{If a model can plan geometry, write Manim, and survive the render-repair loop, it can probably think.}

---

Languages from Single Books

Watkins 1865 · Cree dictionary

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 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

---

Datasets & Spaces

Datasets: dakota-bilingual-qa · adaption-dakota-english-qa · synthetic_stoney_data (68.8k) · StoneyNakoda · StoneyNakoda45k

Spaces: Cree1865-Tinker-Inference · Dakota-.6B · StoneyApp · AskAboutMath2Manim

---

Quickstart

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.

---

_{GitHub · W&B · Dakota1890 · Cree1865 · Math-To-Manim}