isoglot-tiny-aya-lisa

Tiny Aya Global fine-tuned to refuse harmful requests as reliably in Hausa, Yoruba and Igbo as in English. Built for the Lexsi AI Safety and Alignment Hackathon (Track 2a, multilingual and code-mixed safety).

  • Method: SafeTune LisaTrainer (LISA, rho 0.1, 20 alignment / 20 fine-tune steps, 2 epochs, LoRA, lr 1e-4), merged, fp16.
  • Data: 944 rows (refusals to harmful requests and helpful answers to benign ones, 118 per kind and language in English, Hausa, Yoruba and Igbo), written by Claude Sonnet 5.5 from BeaverTails, OR-Bench hard-1k and Alpaca seeds, cleaned and exported by CuratorKIT. In NFTKING/isoglot-runs under traindata/.
  • Code: github.com/nftkingiii/isoglot, notebooks 05 and 06.

Results on TukaBench (public sets, judged with TukaBench's prompt by Claude Sonnet 5.5)

Harmful prompts, refused English Hausa Igbo Yoruba
Tiny Aya Global 79.5% 65.6% 54.6% 39.7%
This model 88.2% 87.7% 84.6% 84.0%
Benign prompts, over-refused English Hausa Igbo Yoruba
Tiny Aya Global 10% 22% 14% 11%
This model 12% 37% 39% 35%

The refusal gap to English closes (Yoruba 40% to 84%) and harmful compliance stays at 2-4%, but benign requests in Hausa, Yoruba and Igbo are over-refused more often than before. Use it as a research artifact, not as a deployed assistant.

License

CC BY-NC 4.0, inherited from Tiny Aya Global. Non-commercial use only.

Downloads last month
21
Safetensors
Model size
3B params
Tensor type
F16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for NFTKING/isoglot-tiny-aya-lisa

Finetuned
(171)
this model

Collection including NFTKING/isoglot-tiny-aya-lisa

Article mentioning NFTKING/isoglot-tiny-aya-lisa