batwan-tiny-aya

A Hiligaynon (Ilonggo) chat model: a LoRA fine-tune of CohereLabs/tiny-aya-water, merged and quantised to GGUF (Q4_K_M) for llama.cpp.

batwan-tiny-aya 1.2.0, trained 2026-11-02 on Hugging Face Jobs (a10g-large, 4 h 6 min, 1,752 steps, held-out loss 0.310), on almost twice 1.0.0's data and with shorter answers in it.

It speaks Hiligaynon. It does not know the dictionary: asked what an unfamiliar word means, it answers fluently and is usually wrong. The Batwan service answers word questions from a word list instead of from this model for that reason. Do not use the weights alone for translation.

Base model	CohereLabs/tiny-aya-water (3.35B parameters)
Method	LoRA, r=16, all linear layers, 2 epochs (scripts/train_batwan.py)
Data	alsersugasawa/batwan-hiligaynon at revision c721c60, 15,573 rows from eleven sources; sources, licences and references in SOURCES.md
File	batwan-tiny-aya-1.2.0-Q4_K_M.gguf in alsersugasawa/batwan-tiny-aya
SHA-256	9e0fe87c2b20ee04c3765a485d20d36366aee18f4ba03bdbd5726e63aaf7d83f
Adapter	alsersugasawa/batwan-tiny-aya-lora-1.2.0; merged weights alsersugasawa/batwan-tiny-aya-merged-1.2.0
Code	batwan-ai-tiny-aya-water v1.2.0
Licence	CC-BY-NC-4.0, inherited from tiny-aya-water and from the training data (PLTAT/alpaca_hiligaynon); some sources are share-alike, and two state no licence (SOURCES.md)
Prompt	src/batwan/prompts.py - the model expects exactly this system prompt, as a system turn

Assessment

scripts/eval_batwan.py, 59 questions, asked of the model directly through llama.cpp (reports in docs/eval/).

	Stock tiny-aya-water Q4_K_M	batwan-tiny-aya 1.0.0	batwan-tiny-aya 1.2.0
Replies in Hiligaynon	0/10	8/10	8/10
Hiligaynon whatever is asked (English, Tagalog)	0/6	5/6	4/6
Word → English (held out)	2/20	2/20	2/20
English → word (held out)	0/20	0/20	0/20
Identity	0/3	3/3	3/3

1.2.0's one miss more is a short reply in Hiligaynon that the test finds too short to call. Twenty everyday questions, three seeds each, 160 tokens, with llama.cpp's DRY (0.8) and a repeat penalty of 1.1 at temperature 0.4:

60 replies	1.0.0	1.2.0
Looping	0 (13 without the penalties)	0 (14 without)
Cut off by the token limit	21	7
Not Hiligaynon	2	0
Median length, characters	394	267

Serve it with those penalties: without them it loops as 1.0.0 did.

Limits

It is a small model taught a low-resource language from about fifteen thousand examples, a large share of them machine-translated. It will make mistakes in grammar and vocabulary, and it can state things that are not true, and often does. Without a repetition penalty a long open answer can repeat itself until it reaches the token limit. Its replies have not been reviewed by a Hiligaynon speaker. Not for commercial use.