Cree1865 Tinker RL Adapter

This model card tracks the Cree1865 reinforcement-learning run built from Rev. E. A. Watkins' 1865 A Dictionary of the Cree Language. The repository is a Dakota1890-derived pipeline replayed on a new historical source.

This is a research artifact and a bootstrap pass. It is not a Cree language authority, a production translator, or a substitute for community review.

Source

• Source: Watkins, E. A. (1865), A Dictionary of the Cree Language
• Local PDF: CreeDictionary.pdf
• Archive master: sources/CreeDictionary_1865_cihm_41985_complete.pdf
• Structure: one book with Part I. English-Cree and Part II. Cree-English

Dataset Snapshot

Latest local full-dictionary build:

Field	Value
Extracted page JSON files	463
Raw entries	19,607
Deduplicated usable entries	19,560
Rejected incomplete entries	125
Multi-variant entries	4,049
SFT train / valid	18,463 / 972
RL task records	38,870
Plain Q&A records	38,870
English->Cree RL tasks	19,435
Cree->English RL tasks	19,435
Duplicate RL task IDs after normalization	0

The generated data lives under data/cree_goal_run_20260624_full_dictionary/ locally and is ignored by git. The extraction and dataset-building code is tracked so the artifacts can be regenerated.

Current Training Status

Redesigned full-dictionary showcase run in progress:

• Launch date: 2026-06-25
• Base model: Qwen/Qwen3-30B-A3B-Instruct-2507
• Dataset: data/cree_goal_run_20260624_full_dictionary/training_datasets/rl_tasks_balanced_cree_showcase.jsonl
• Weighted training rows: 70,040
• Original rows covered: 38,870 / 38,870
• Direction balance: 35,020 English->Cree / 35,020 Cree->English
• Objective: grouped rollout RL with Tinker importance_sampling
• Batch size / group size: 32 / 8
• Planned steps: 2,189
• LoRA rank: 64
• W&B run: 4om7k9ao

Preliminary held-out eval snapshot from step 500:

Eval metric	Step 0	Step 500	Delta
Overall verifier reward	0.2358	0.3628	+0.1270
Target-Cree reward	0.2835	0.4233	+0.1398
Target-English reward	0.1322	0.2316	+0.0994
Target-Cree exact match	0.0029	0.0485	+0.0456
Target-Cree containment	0.0913	0.1441	+0.0528
Target-Cree orthography	0.4177	0.5400	+0.1222
Target-Cree character F1	0.3504	0.6020	+0.2515
Target-English character F1	0.2818	0.4003	+0.1185

These numbers are preliminary verifier metrics, not evidence of fluency or community validation. The scalar reward is rubric-shaped; the more important signal is per-channel movement and the direction split between English->Cree and Cree->English.

Completed 1200-step run:

• Base model: Qwen/Qwen3.5-4B
• Dataset: data/cree_goal_run_20260624_full_dictionary/training_datasets/rl_tasks_all.jsonl
• Renderer: qwen3_5_disable_thinking
• Steps: 1200
• Batch size / group size: 2 / 2
• Max sampled tokens: 64
• W&B run: kjn02ee4
• W&B project: cree1865-tinker
• Final reward: 0.21
• Deduped mean reward: 0.18260238803447346
• Final parse success: 1.0
• Deduped mean parse success: 0.99875
• Tinker weights: tinker://bf25e2aa-6b3a-557c-8133-fadf5ebcba8f:train:0/weights/final
• Tinker sampler weights: tinker://bf25e2aa-6b3a-557c-8133-fadf5ebcba8f:train:0/sampler_weights/final
• Raw ledger: wandb_analysis/cree_reward_ledger_tinker_full_dictionary_1200step_20260624_qwen35_4b_no_think.csv
• Deduped ledger: wandb_analysis/cree_reward_ledger_tinker_full_dictionary_1200step_20260624_qwen35_4b_no_think_deduped.csv

Resume note:

The first Tinker session stalled at local step 868 with No progress made in 7200s. The run was resumed under the same W&B run ID, kjn02ee4, from checkpoint 000800. The final local raw ledger therefore has 1269 rows: 1200 unique steps plus 69 replay rows from steps 800-868. The deduped ledger keeps the last row per step for steps 0-1199.

Completed smoke run:

• Steps: 4
• Final reward: 0.1894479405034325
• Final parse success: 1.0
• Tinker weights: tinker://096ba4d7-bccc-5d33-9209-e1a1a8d746dc:train:0/weights/final
• Ledger: wandb_analysis/cree_reward_ledger_tinker_full_dictionary_smoke_20260624_qwen35_4b_no_think.csv

At these small-model settings, 1200 steps is long but feasible; the practical risk is Tinker service stalling, not the step count itself. Future longer runs should use checkpointed chunks and preserve a single W&B run ID when continuing the same experiment.

Reward Surface

The current Prime/Verifiers environment is Cree-specific: harleycooper/cree1865-dictionary-qa version 0.1.2. It uses the Watkins dictionary QA export when repo data is present and ships a tiny packaged smoke dataset so Prime registry CI can load the environment in isolation.

The Cree task export provides Tinker-compatible fields:

• question
• prompt
• answer
• task_type
• difficulty
• info.special_chars
• info.verification_pattern
• metadata.direction

Reward channels are exact match, target containment, Cree orthography preservation, character F1, and concise answer length. There is no Dakota grammar rubric and no affix/default reward channel in the Cree verifier. Larger runs should still review long reverse-section English glosses and community-facing Q&A wording before making model-quality claims.

Limitations

• The source is a missionary-era 1865 dictionary with historical orthography and colonial-era framing.
• The extraction may preserve source errors, scan artifacts, and model extraction mistakes.
• Many tasks are direct dictionary lookups, not natural dialogue.
• Long reverse-section glosses should be filtered or down-weighted before larger-scale training.
• Cree language work should be reviewed with appropriate community and linguistic expertise.

Citation

Watkins, E. A. (1865). A Dictionary of the Cree Language, as Spoken by the Indians of the Hudson's Bay Territories. London: Society for Promoting Christian Knowledge.