Skip to content

MODEL_CARD.showcase.md

Latest commit

 

History

History
147 lines (118 loc) · 6.11 KB

MODEL_CARD.showcase.md

File metadata and controls

147 lines (118 loc) · 6.11 KB
Raw
language
crk
en
license apache-2.0
library_name peft
pipeline_tag text-generation
base_model Qwen/Qwen3.5-4B
tags
cree
indigenous-languages
low-resource-language
dictionary
reinforcement-learning
grpo
tinker
thinking-machines
peft
lora
datasets
local:cree_goal_run_20260624_full_dictionary
widget
text
Translate 'a good man' to Cree, preserving the 1865 orthography. Return only the answer.
text
Give the English meaning of the Cree word Pāsooch. Return only the answer.

Cree1865

A 1200-step GRPO/Tinker bootstrap run from Rev. E. A. Watkins' 1865 Cree dictionary.

Overview

Cree1865 is a research bootstrap artifact built from a single public-domain source: Rev. E. A. Watkins' 1865 A Dictionary of the Cree Language. The repo replays the Dakota1890 extraction-to-RL pipeline on Cree, producing structured dictionary entries, SFT/Q&A pairs, and deterministic reward-ledger RL tasks.

This is not a Cree language authority, a production translator, or a substitute for community review. It is a transparent technical starting point for inspection and correction.

Source And Dataset

Field Value
Source Watkins 1865, A Dictionary of the Cree Language
Internet Archive ID cihm_41985
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 / validation 18,463 / 972
RL task records 38,870
Plain Q&A records 38,870
English-to-Cree RL tasks 19,435
Cree-to-English RL tasks 19,435

The local dataset root is data/cree_goal_run_20260624_full_dictionary/. Generated data is ignored by git; the extraction and dataset-building code is tracked in the repository.

Training Run

Field Value
Base model Qwen/Qwen3.5-4B
Renderer qwen3_5_disable_thinking
Method GRPO with deterministic reward ledger
Steps 1200
Batch size / group size 2 / 2
Max sampled tokens 64
W&B project cree1865-tinker
W&B run kjn02ee4
Final reward 0.21
Deduped mean reward 0.18260238803447346
Final parse success 1.0
Deduped mean parse success 0.99875
Final Tinker weights tinker://bf25e2aa-6b3a-557c-8133-fadf5ebcba8f:train:0/weights/final
Final sampler weights tinker://bf25e2aa-6b3a-557c-8133-fadf5ebcba8f:train:0/sampler_weights/final

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

Reward Ledger

Tracked ledgers:

  • Raw: wandb_analysis/cree_reward_ledger_tinker_full_dictionary_1200step_20260624_qwen35_4b_no_think.csv
  • Deduped: wandb_analysis/cree_reward_ledger_tinker_full_dictionary_1200step_20260624_qwen35_4b_no_think_deduped.csv

The published Cree verifier is harleycooper/cree1865-dictionary-qa version 0.1.2. 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. The verifier is deterministic; there is no LLM judge in the reward path.

Usage

The final Tinker checkpoint is recorded above. Once an adapter is published, it should load with PEFT on top of Qwen/Qwen3.5-4B.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_name = "Qwen/Qwen3.5-4B"
adapter_name = "HarleyCooper/Cree1865"  # replace with the published adapter id

model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
model = PeftModel.from_pretrained(model, adapter_name)

messages = [
    {"role": "system", "content": "You are a Cree language assistant. Return only the answer."},
    {"role": "user", "content": "Translate 'a good man' to Cree, preserving the 1865 orthography."},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=64, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

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.
  • The current task set is primarily direct dictionary lookup Q&A, not conversational Cree competence.
  • Long reverse-section English glosses should be reviewed or filtered before larger runs.
  • 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. Internet Archive: cihm_41985.

@misc{cree1865_model,
  title  = {Cree1865: A GRPO Cree Language Adapter from a Single 1865 Dictionary},
  author = {Cooper, Christian Harley},
  year   = {2026},
  note   = {Base: Qwen/Qwen3.5-4B. Source: Watkins 1865, IA cihm_41985.
            Method derived from Dakota1890.}
}