Wolof Qwen2.5-1.5B — conversational + translation (wo · fr · en)
A fine-tune of Qwen2.5-1.5B-Instruct for the Wolof language. It can hold a chat conversation and translate between Wolof ↔ French ↔ English.
Trained with Unsloth (QLoRA, 4-bit) on a single GPU.
Intended use
- Wolof ⇄ French ⇄ English translation inside a chat interface
- Experiments and research on low-resource (Wolof) language modelling
- A base to iterate on with more conversational Wolof data
How to use
Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("ciskoM/wolof-qwen-1.5b")
model = AutoModelForCausalLM.from_pretrained("ciskoM/wolof-qwen-1.5b", device_map="auto")
msgs = [{"role": "user", "content": "Translate to Wolof: Good morning, how are you?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=64)[0], skip_special_tokens=True))
GGUF (Ollama / LM Studio / llama.cpp) — see the companion repo
ciskoM/wolof-qwen-1.5b-gguf (q4_k_m). With Ollama:
ollama create wolof -f Modelfile # Modelfile: FROM ./wolof-qwen.Q4_K_M.gguf
ollama run wolof
Example prompts: Translate to Wolof: …, Traduis en wolof : …,
What does "Jamm rekk" mean in English?, Nanga def?
Training data
~120k instruction pairs (ShareGPT format), balanced across translation directions plus monolingual Wolof:
| Source | Content |
|---|---|
bilalfaye/english-wolof-french-translation | aligned en/wo/fr sentences |
galsenai/centralized_wolof_french_translation_data | aligned wo/fr sentences |
| ALMA / DLIR Wolof e-books (13) | OCR'd authentic Wolof prose |
| wolofresources.org PDFs + UDHR Wolof | proverbs, tales, vocabulary |
| Wolof New Testament (bibliamundi) | clean verse-numbered scripture |
The dataset is ~96% translation pairs; bidirectional pairs (en2wo, wo2en, fr2wo, wo2fr) were generated with varied prompt templates.
Training procedure
- Base:
Qwen/Qwen2.5-1.5B-Instruct, loaded 4-bit - LoRA: r=16, alpha=16, dropout=0, on all attention + MLP projections
- 1 epoch, lr 2e-4, linear schedule, adamw_8bit, seq len 2048
- Trained on assistant responses only (
train_on_responses_only) - Tooling: Unsloth + TRL
SFTTrainer/SFTConfig
Limitations & honest caveats
- Translation > conversation. Because the data is overwhelmingly translation pairs, the model is strong at translating but limited at free-flowing Wolof conversation. More conversational Wolof data is the fix, not a bigger model.
- Wolof is low-resource; expect errors, especially on rare topics, long inputs, and Wolof orthography variants.
- Some training text came from OCR (scanned PDFs) and subtitle-sourced corpora — there may be noise, odd register, or artifacts.
- Not safety-tuned; do not use for high-stakes decisions.
Data & licensing note
Model weights derive from Qwen2.5 (Apache-2.0). However, the training data provenance is mixed: the translation corpora are subtitle/film-sourced and the scripture is from a translation that may carry its own copyright. This card is released CC-BY-NC-4.0 as a conservative default. Verify the licenses of the underlying datasets before any commercial use or redistribution.
Acknowledgements
Datasets by bilalfaye and the GalsenAI community; ALMA/DLIR African Language Materials Archive; wolofresources.org; Biblia Mundi. Built with Unsloth.