catinthebag/Gatra-1-JawaKrama

GatraOne (Gatra-1) is a synthethic Jawa Krama instruction-tuning dataset, generated by GPT-4.

Introducing the Gatra-1 dataset

This is a synthetic dataset to fine-tune LLMs into responding in Jawa Krama, the high-register of Javanese language. It is 98% generated using GPT-4, which has very good Jawa Krama capabilities. It is currently a 'beta' version with only 560 input-output prompts.

So far, this has been only tested on fine-tuning GPT-3.5 with considerable success. The idea behind this fine-tuning is that we noticed that several LLMs, including but not limited to GPT-3.5, is incapable of responding in Jawa Krama. However, certain models appear to 'understand' Jawa Krama text to some degree. To test this one can copy and paste a text in Jawa Krama to a model and ask them what the text is about.

The case with a lot of major models will nevertheless understand what the text is about, although sometimes with hallucinations. This is also what we notice when working with GPT-3.5. We hypothesized that the model's latent linguistic understanding can be 'brought out' through fine-tuning. We are very pleased that we were succesful when testing with fine-tuning GPT-3.5 (note that GPT-3.5 can respond in Jawa Ngoko, although sometimes with noticeable mistakes).

As one can easily imagine, this provides tremendous potential for low-resource language preservation. It is likely that major models contain bits and pieces of low-resource languages in their training data, but they were not further tuned to respond in them. We find that even small models like Mistral 7B possess understanding of Jawa Krama text. This shows that low-resource language preservation (at least, for certain languages) may not need considerable pre-training data, and that bits and pieces embedded in major models' training data can be awakened through instruction tuning.

There are still a great deal of room for improvement, but we are very excited with future prospects! We hope that this can contribute to the development of low-resource language technologies, especially through fine-tuning LLMs.

Dataset details

The dataset comprises 273654 tokens based on OpenAI's GPT-4 tokenizer, including the system instructions. The data is almost entirely synthetic with minimal human curation. Most of the curation we do lie in fixing linguistic errors by GPT-4 when it used a Ngoko word when a distinctly Krama one is available. Only in some cases, we modify or even create ourselves the models' output.

The majority of the dataset comprises of NLG with respect to education. Mostly we create QA pairs about academic disciplines. For NLU cases (currently, all of them are text comprehension), the texts themselves are also synthetically generated.

One method that we do to significantly improve question creation time is that we ask the model to create QA pairs instead of answering our questions. For example, one prompting strategy is prompting "Please create a QA pair about x" where x is modified with the name of an academic discipline. With this simple means we are capable of creating hundreds of QA pairs with some degree of depth to the questions. So, although we are not physicist, we are capable to nevertheless create a QA pair about physics in some depth.

Note that GPT-4 typically uses inconsistent Krama, and a fair amount of prompt engineering is needed to ensure it is constrained to respond in Krama.

As mentioned, this dataset is mostly focused on educational/academic QA. We hope to expand the proportion of domains in the future.

Acknowledgments

This dataset is authored by Afrizal Hasbi Azizy, though we prefer to use the royal we.

We receive partial financial support from dflabs.id for the creation of this dataset. Matur nuwun!

---

Interaction Examples

Below are examples of GPT-3.5-0125 responding in Jawa Krama after being fine-tuned with our dataset. Note that the model made mistakes every now and again, but it managed to be mostly-correct. In this examples we use the original system instruction as included in the dataset. The fine-tuning cost for this full dataset were ~6.7 USD.

Prompt Menapa kodhok mboten saged nafas ing jero toya?

FT GPT-3.5-0125 Kodhok mboten saged nafas ing jero toya amargi kodhok kagungan sistem respirasi ingkang beda kados ta kanthi manungsa lan satwa sanesipun. Satwa kodhok ngangge kulit minangka alat respirasi ingkang wigatos. Kulit kodhok kagungan kapasitas kangge ngaliraken oksigen lan karbon dioksida kanthi efisien. Kulit kodhok kagungan pembuluh darah ingkang sae sisih ingkang kapasitasipun gede, ingkang ngidinaken pertukaran gas ingkang sae. Kanthi cara punika, kodhok saged nafas ing jero toya lan tetep ngidup tanpa kudu nafas lewat paru-paru kados ta manungsa.

Sistem respirasi kodhok ingkang unik punika ngidinaken kodhok saged adaptasi kaliyan lingkunganipun ingkang beda-beda. Kodhok saged ngalami proses respirasi ingkang efisien sajeroning lingkungan ingkang oksigenipun asring, kados ta toya. Kanthi ngalami proses respirasi ingkang unik punika, kodhok saged ngidup lan berkembang biak ing lingkungan ingkang sanesipun. Kulit kodhok kagungan lapisan tipis ingkang ngidinaken pertukaran gas ingkang sae, lan sistem sirkulasi darah ing kulitipun wigatos kangge ngaliraken oksigen lan nutrisi dhateng seluruh badan kodhok.

---

Prompt plato ngendikakaken punapa babagan sastra?

FT GPT-3.5-0125 Plato minangka salah satunggaling filosof Yunani ageng ingkang gadhah pengaruh ageng dhumateng perkembangan sastra. Ing kawicakanipun, Plato nyritakaken babagan sastra dhumateng dialogipun ingkang asring dipunwastani dening Socrates minangka karakter utami. Plato nyataken sastra minangka imajinasi kedah, lanipun mboten dados sumber pituduh ing babagan kados ta. Plato ngajeni sastra ingkang dipunwastani ing Yunani Kuno, kados ta tragedi lan epik, nangingipun ngajeni sastra ingkang nggadhah kritik dhumateng kebenaran lan keadilan. Punika, sastra ingkang dipunanggep bermutu saged mitulung individu nggayuh kearifan lan kebenaran, nanging sastra ingkang mboten bermutu saged dados pengaruhipun ingkang mligi lan nggadhahi opini ingkang mboten bener.

Salajengipun, Plato ugi nyataken babagan konsep mimesis, utawi tiruan, ing sastra. Punika nggambarkan carane sastra nggambarkaken alam semesta lan realitas ing sekitaripun. Plato nyataken babagan mimesis ing sastra kados ta tragedi lan komedi minangka bentuk imitasi ingkang langkung rendah, amargi mboten ngajeni kebenaran lan keadilan. Punika, Plato mawon nyataken babagan babagan sastra ingkang ngupayakaken ngajeni kebenaran lan keadilan, kados ta dialog, minangka bentuk sastra ingkang langkung dianggep bermutu lan saged mitulung individu nggayuh kearifan.

---

Prompt aku wis ra ngerti kudu piye meneh, saben dina aku kangen karo bojoku

FT GPT-3.5-0125 Kula ngertos perasaan sami dalem ingkang dipunrasakaken, utawi dhumateng sanesipun. Sanadyan panjenengan kedah nggayuh pundi ingkang wigatos, dene kados pundi cara-cara ingkang saged dipunlakuaken supados perasaan panjenengan saged dipunsebat utawi dipunkawisaken. Panjenengan saged nglakuaken hal-hal ingkang saged nyiptakaken kedah rasa kawigatenipun, kados ta ngajengaken komunikasi lan kasil saking jauh-jauhan. Kajawi punika, panjenengan saged ngrembakaaken pikiripun dhumateng hal-hal sanesipun ingkang saged nyenengaken lan nyegah perasaan kawigatenipun. Ing babagan punika, panjenengan saged nggadhah diri kados ta supados saged nglakuaken kegiatan utawi hobbi ingkang panjenengan senengaken, ngajengaken kawicaksanan, lan mbudidaya diri dhumateng kados pundi aspek ingkang saged dipunkontrol, supados perasaan kawigatenipun saged dipunatasi.