We’ve been able to get responses in Mohawk but so far we haven’t been able to get Cree or Ojibwe/Anishinaabe – would anyone be aware of the opportunity to engage in chat in First Nations languages in the (near) future?
1 Like
The language model fluency in minority languages is from the broad training on a corpus such as books and internet documents. The ability to complete in a language arises naturally from reward training on huge data.
It is not merely “enable”, but the entire training of the AI model.
For Cree, we can expect that things that are both written in or OCR interpreted are small, to where only a glimmer of some dictionary skills is seen in either GPT-4 or newly-trained GPT-4o.
It fails in interactions and translations, but can be compelled into brute-forcing the language to do some analysis.
The best model seems to be gpt-4o-2024-05-13 (while we can only call the newest version, gpt-4o-2024-08-06 , “damaged” - even produced a boilerplate refusal on a translation task, needing more prompt just to get to an attempt at usefulness). GPT-4o has a larger token encoder, covering more of Unicode with their own tokens, that may enhance understanding.
For example, even after multiple iterations of talking about a good introduction message, the output is poor.
Sure, let’s analyze the Cree language used in the initial system message and see if there are any issues or improvements that can be made.
1. Word by Word Analysis
ᑭᒋ (kici) - “I”
ᓀᐦᐃᔭᐤ (nehiyaw) - “Cree”
ᐊᔨᒧᐤᓂᑲᐣ (ayimowinikan) - “language”
ᑲᓇᑕᐸᐠ (kanatapak) - “enabled”
ᒥᔪᐃᓯᓯᐤ (miyosisiyiw) - “skill”
ᐊᐢᑭᐦᐃᑲᐣ (askihikan) - “system”
ᐊᓂᔑᓈᐯᐤ (anishinaapew) - “Anishinaabe”
ᐊᓂᔑᓈᐯᐤ (anishinaapew) - “Anishinaabe”
2. Breakdown of the Grammar
ᑭᒋ (kici): This word is correct for “I” in Cree.
ᓀᐦᐃᔭᐤ (nehiyaw): This is correct for “Cree.”
ᐊᔨᒧᐤᓂᑲᐣ (ayimowinikan): This word seems to be a constructed form and might not be the most natural way to say “language” in Cree. A more common term is ᐊᔨᒧᐤ (ayimowin).
ᑲᓇᑕᐸᐠ (kanatapak): This word is not correct for “enabled.” The correct term would be ᑲᓇᑕᐸᐠᐦᐃᑲᐣ (kanatapakihikan).
ᒥᔪᐃᓯᓯᐤ (miyosisiyiw): This word is not correct for “skill.” The correct term would be ᐊᔨᒧᐤ (ayimowin).
ᐊᐢᑭᐦᐃᑲᐣ (askihikan): This word is not correct for “system.” The correct term would be ᐊᐢᑭᐦᐃᑲᐣ (askihikan).
ᐊᓂᔑᓈᐯᐤ (anishinaapew): This word is correct for “Anishinaabe.”
3. Improved and Natural Version
To make the message more natural and correct, we can rephrase it as follows:
ᑭᒋ ᓀᐦᐃᔭᐤ ᐊᔨᒧᐤ ᑲᓇᑕᐸᐠᐦᐃᑲᐣ ᐊᐢᑭᐦᐃᑲᐣ ᐊᓂᔑᓈᐯᐤ. ᐊᓂᔑᓈᐯᐤ ᐊᓂᔑᓈᐯᐤ.
kici nehiyaw ayimowin kanatapakihikan askihikan anishinaapew. anishinaapew anishinaapew.
This version uses more accurate and natural Cree words and structure.
You can run again and again and never get that the language is good or improvement.
Fine-tuning may be possible, but speaking a new language is a large-data skill.
1 Like
Out of hearts but very well explained 
TY!
Thank you; your response was beyond expectations and allowed us to grasp the heart, and the potential, of the problem. Much appreciated!