Iamjuhwan/HY-MT1.5-yoruba-pidgin-lora

from huggingface_hub import upload_file model_card = """--- language: - yo - en - pcm base_model: tencent/HY-MT1.5-1.8B tags: - translation - lora - peft - yoruba - nigerian-pidgin - african-languages - qlora license: apache-2.0 --- # HY-MT1.5-1.8B — Yoruba & Nigerian Pidgin LoRA Fine-tuned version of tencent/HY-MT1.5-1.8B on Yoruba-English and Nigerian Pidgin-English translation pairs using QLoRA. ## Evaluation Results (Baseline vs Fine-Tuned) Evaluated on 299 clean Yoruba/Pidgin/English examples, filtered from noisy OPUS-100 software strings: | Metric | Base Model | Fine-Tuned | Delta | |--------|------------|------------|-------| | BLEU | 6.21 | 6.90 | +0.70 | | chrF | 13.57 | 13.59 | +0.03 | BLEU scores are characteristically low for morphologically rich low-resource languages like Yoruba. The +0.70 BLEU improvement is meaningful given the small training set (~2,000 pairs). chrF improvement was marginal, indicating that larger, higher-quality agricultural domain data would produce stronger gains. ## Motivation The base model was evaluated across 12 adversarial probe categories (see Jesujuwon/HY-MT1.5-1.8B-blindspots). Key failure modes identified: - Literal translation of idiomatic expressions - Hallucination on low-resource African language inputs - Poor handling of code-switched Pidgin/Yoruba/English text - Loss of sarcastic and ironic register ## Training Data | Source | Pairs | Languages | |--------|-------|-----------| | OPUS-100 (Helsinki-NLP) | ~4,000 | English and Yoruba | | Jesujuwon/HY-MT1.5-1.8B-blindspots | 12 | Multi (correction pairs) | | Custom synthetic pairs | 24 | Nigerian Pidgin and English, Yoruba and English | ## Training Details | Parameter | Value | |-----------|-------| | Base model | tencent/HY-MT1.5-1.8B | | Method | QLoRA (4-bit NF4 + LoRA) | | LoRA rank | 16 | | LoRA alpha | 32 | | Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj | | Epochs | 3 | | Effective batch s …