---
base_model: cis-lmu/glot500-base
language:
- crk
license: cc-by-4.0
pipeline_tag: text-classification
tags:
- figurative-language
- text-classification
- plains-cree
- clkd
- low-resource
---

# glot500-base-plains-cree-en-clkd-direct

cis-lmu/glot500-base adapted for Plains Cree figurative language detection via Cross-Lingual Knowledge Distillation (CLKD) (no TLM warmup).

## Method

Inspired by [Cross-Lingual Knowledge Distillation (ACL 2023)](https://aclanthology.org/2023.findings-acl.885/).

**CLKD pipeline:**
1. Teacher: `KonradBRG/deberta-v3-base-figurative` (frozen, English)
2. Student base: `cis-lmu/glot500-base`
3. All layers trainable.

The teacher predicts soft label distributions on English translations of Plains Cree sentences.
The student is trained to match these distributions on the Cree side via KL divergence.

**Training data:** ~9,000 Cree–English parallel sentence pairs
**Temperature:** 2.0 | **Epochs:** 10 | **Hardware:** 1× NVIDIA A100 40 GB

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


## Citation

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


## Data

Training data includes:
- [Bloomfield (1934) *Plains Cree Texts*](https://bloomfield.kiyanaw.net) (scraped and sentence-aligned)
- [EdTeKLA Indigenous Languages Corpora](https://github.com/EdTeKLA/IndigenousLanguages_Corpora)