Files
Table with columns: Path, Stage, Tinker checkpoint| Path | Stage | Tinker checkpoint |
|---|
| repository root | supervised warm start, then 30 RL steps | tinker://516a9067-2fe6-5624-a6a7-5b9f07e84d8e:train:0/weights/final |
stage1-supervised/ | supervised warm start only | tinker://2897218c-c3ef-5647-8215-65a4e745b7a7:train:0/weights/final |
Training
Table with columns: Stage, Data, Size, Cost| Stage | Data | Size | Cost |
|---|
| 1. Supervised | 56,120 chat rows, six single-turn task types | one epoch, 436 steps of 128 rows, learning rate 2e-4 | $3.61 |
| 2. RL | 68,911 tasks, eight task types, deterministic rewards | 30 steps of 32 tasks × 8 rollouts, learning rate 2e-5, temperature 1.0 | $1.12 |
The RL loss is the importance-sampling policy gradient of the cookbook, with advantages centred inside each group of rollouts. There is no LLM judge.
Evaluation
Exact-match rate, 40 tasks per task type, one seed, sampling temperature 1.0.
Table with columns: Task, Split, Base model, stage1-supervised/, root (supervised + RL)| Task | Split | Base model | stage1-supervised/ | root (supervised + RL) |
|---|
syllabify | held-out word families | 0.00 | 0.83 | 0.83 |
pos | held-out word families | 0.05 | 0.38 | 0.73 |
root | held-out word families | 0.00 | 0.48 | 0.58 |
inflect | held-out word families | 0.00 | 0.30 | 0.38 |
lookup_gloss | held-out word families | 0.50 | 0.00 | 0.85 |
lookup_root_gloss | held-out word families | 0.10 | 0.00 | 0.63 |
gloss (token F1) | recall | 0.00 | 0.05 | 0.08 |
en_to_dak | recall | 0.00 | 0.00 | 0.03 |
What this shows and does not show:
- The adapters learn rules of the book that transfer to unseen words (syllable division and accent).
- RL restores and improves the dictionary-lookup behaviour that supervised training removed.
- The adapters do not know the Dakota vocabulary. Recall of definitions and of English→Dakota is near zero.
- Differences of a few points between columns are inside the noise of 40 tasks.
The adapters were trained with the qwen3_5_disable_thinking chat format and these system prompts.
Single-turn tasks:
You answer from Stephen R. Riggs, A Dakota-English Dictionary (1890). Use the spelling of that book. Give only the answer on one line, starting with 'ANSWER:'.
Example user message:
Write the Dakota word with Riggs' syllable division and accent mark (example: a-hi'-yo-ka-sin): ahiyokasin
The lookup tasks need the dictionary tool of the environment. See the environment README in the repository.
Use
The files are a PEFT LoRA adapter, converted from the Tinker checkpoint with tinker_cookbook.weights.build_lora_adapter. The conversion targets serving with vLLM or SGLang:
vllm serve Qwen/Qwen3.5-4B --enable-lora --lora-modules riggs=HarleyCooper/Qwen3.5-4B-Riggs1890-lexgym-pilot
We evaluated the checkpoints on Tinker. We have not yet load-tested the converted adapter files in vLLM or in peft.
Limits
- The source is a missionary dictionary from 1890. It records one written variety and is not a description of Dakota as it is spoken now.
- The training data comes from a vision-model reading of page scans and contains reading errors.
- Output of this model is not authoritative Dakota and carries no community endorsement. A model that is correct for the 1890 book is a starting point for correction by Dakota speakers.