xlm-mlm-abl-no-tlm-plains-cree-en-calibrated

FacebookAI/xlm-mlm-100-1280 calibrated on DeepSeek-annotated Bloomfield Plains Cree sentences for figurative language detection.

Training

Low-LR calibration of a CLKD checkpoint on gold/silver Bloomfield validation data (all DeepSeek-annotated sentences).

Ablation condition: no TLM warmup — base model → CLKD → Calibrate

Base checkpoint: FacebookAI/xlm-mlm-100-1280 Literal : figurative ratio: 3:1 | Epochs: 15 | Learning rate: 5e-06 Hardware: 1× NVIDIA A100 40 GB

Labels: literal (0), idiom (1), metaphor (2), simile (3)

Intended use

Final deployed checkpoint for figurative language detection in Plains Cree.

Citation

If you use this model, please cite the associated thesis/paper (TBD).

Data

Training data includes:

• Bloomfield (1934) Plains Cree Texts (scraped and sentence-aligned)
• EdTeKLA Indigenous Languages Corpora