Qwen3-0.6B Kalenjin
An experimental full fine-tune of Qwen3-0.6B for bidirectional English ↔ Kalenjin translation.
This project explores whether a small general-purpose language model can be adapted for low-resource machine translation, rather than using a dedicated translation architecture such as NLLB or Marian.
Model
- Base model:
Qwen/Qwen3-0.6B
- Training: Full-parameter fine-tuning
- Dataset:
mutaician/english-kalenjin-dataset
- Clean training pairs: 50,768
- Bidirectional training examples: 101,530
- Maximum sequence length: 256
- Effective batch size: 128
- Learning rate: 2e-5
- Selected checkpoint: Step 2200
- Best validation loss: 1.8685
The selected checkpoint was chosen based on validation performance. Later checkpoints achieved substantially lower training loss but showed clear overfitting and poorer qualitative performance.
Evaluation
Evaluation was performed on a strict held-out test set of 5,635 sentence pairs. English source sentences occurring in the cleaned training split were excluded from this test set.
| Direction | BLEU | chrF | chrF++ |
|---|
| English → Kalenjin | 5.98 | 35.87 | 31.13 |
| Kalenjin → English | 20.86 | 40.65 | 39.28 |
These scores should be interpreted cautiously because Kalenjin is a low-resource, morphologically rich language and the test set contains a single reference translation per example.
Usage
The model was trained using explicit translation instructions.
English → Kalenjin
1Translate the following English text into Kalenjin. Return only the translation.
2
3Where are you going today?
Example output:
Kalenjin → English
1Translate the following Kalenjin text into English. Return only the translation.
2
3Iwendi ano raini?
Limitations
This model is an experimental research model trained primarily for English ↔ Kalenjin translation.
It was not trained as a general-purpose Kalenjin conversational assistant. Arbitrary prompts outside the translation format may result in repetition, copying of the input, hallucinations, or otherwise unreliable responses.
Translation quality may also vary across dialects, domains, sentence complexity, and terminology.
