Makandal
yon AI ki pale kreyòl, epi ki konn chèche
an AI that speaks Haitian Creole, and knows how to look things up
Yon ti modèl 1 GB ki mache sou pwòp òdinatè w. Li reponn an kreyòl, franse, panyòl ak angle, kout e klè, epi li rele zouti pou lè a, tan an, kalkil ak istwa.
A small 1 GB model that runs on your own computer. It answers in Kreyòl, French, Spanish and English, short and clear, and calls tools for the time, the weather, arithmetic and history.
Ayiti nan epòk AI aHaiti in the age of AI
Kote kreyòl la kanpe
Where Haitian Creole stands
- Apeprè 12 milyon mounAbout 12 million peoplepale kreyòl ayisyenspeak Haitian Creole
- Lang ofisyèl depi 1987Official language since 1987ansanm ak franse, nan Konstitisyon analongside French, in the Constitution
- 29.5% → 39%lekti an kreyòl, anvan ak apre fòmasyonKreyòl reading, before and after training
Modèl AI yo aprann sitou nan tèks angle. Nou bay modèl debaz la, Gemma 3 1B, yon egzamen lekti: Belebele, 200 tèks ak kat chwa chak. Chwazi o aza bay 25%. An angle, modèl la jwenn 47.5%. An kreyòl, 29.5%, prèske o aza. Lè yon modèl pa konprann kreyòl, se pifò Ayisyen ki rete deyò.
AI models learn mostly from English text. We gave the base model, Gemma 3 1B, a reading test: Belebele, 200 passages with four choices each. Guessing scores 25%. In English the model scores 47.5%. In Kreyòl, 29.5%, barely above guessing. When a model does not understand Kreyòl, most Haitians are left out.
Apre fòmasyon an, makandal v2 rive 39% an kreyòl, e distans ak angle a desann de 18 a 11.5 pwen. Se yon premye pa, ak yon modèl ase piti pou l mache kote entènèt la chè oswa li koupe.
After training, makandal v2 reaches 39% in Kreyòl, and the distance to English shrinks from 18 to 11.5 points. It is a first step, with a model small enough to run where the internet is expensive or cuts out.
ZoutiTools
Li pa devine. Li chèche.
It does not guess. It looks it up.
- Tan, lè, kalkilWeather, time, arithmetic
- Ansiklopedi ak rechèchEncyclopedia and search
- Memwa ak konesans pa wMemory and your own knowledge
Yon modèl piti pa ka sonje tout bagay. Kidonk makandal v2 aprann rele yon zouti lè l bezwen yon fè. Kat egzanp sa yo se vrè repons modèl la ak vrè zouti: tan an soti nan Open-Meteo, istwa a soti nan Wikipedya.
A small model cannot remember everything. So makandal v2 learned to call a tool when it needs a fact. These four are the model's real answers with real tools: the weather comes from Open-Meteo, history from Wikipedia.
Tan anWeather
Ki tan l ap fè Okap kounye a?
Nan Okap li fè 26 degre, prèske klè, imidite a 84%.
IstwaHistory
Kiyès Jean-Jacques Dessalines te ye?
Jean-Jacques Dessalines se te yon gwo lidè militè ak chèf Revolisyon Ayisyen an. Li te pwoklame endepandans peyi a 1ye janvye 1804 epi li te vin premye anperè Ayiti.
KalkilArithmetic
Si m achte 3 sak diri a 1250 goud chak, konbyen sa fè?
Sa fè 3750 goud. Se kantite la pou 3 sak diri a.
An franseIn French
Quel temps fait-il à Jacmel ?
Il fait 29 degrés à Jacmel avec un ciel dégagé. L'humidité est de 67%.
Yon konvèsasyon li pa t janm wè anvanA conversation it never saw in training
Mwen bezwen yon lide pou manje... pou fanmi an. Yon bagay ki bon pou yon gwoup moun.
I need a meal idea for the family, something good for a group.
Yon diri ak pwa ap mache byen. Se yon plat tout moun renmen. Èske ou vle m ba ou resèt la?
Rice and beans would work well. Everyone loves it. Want the recipe?
E vyann kochon an, ki jan pou m kwit li pou l pa sèch?
And the pork, how do I cook it so it isn't dry?
Mete yon ti dlo nan chodyè a. Ajoute yon ti sèl ak lay tou. Kite l bouyi dousman pandan 20 minit. Sa ap fè vyann lan mou.
A little water in the pot, salt and garlic, simmer for 20 minutes. That keeps it tender.
Ok, m konprann. Mwen pral al nan mache a kounye a.
OK. I'm going to the market now.
Vyann kochon an ap gou anpil. Pa bliye mete yon ti epis ladan l tou. Bon chans nan mache a.
The pork will be tasty. Don't forget some seasoning. Good luck at the market.
Teste lTest it
Menm kesyon, de modèl
Same question, two models
- Kesyon modèl la pa t janm wèQuestions it never saw
- Repons kopye jan yo sotiAnswers copied as they came out
Chwazi yon kesyon. A goch, Gemma 3 1B anvan fòmasyon an. A dwat, makandal v2. Gade lang lan, longè a, ak si repons lan gen sans.
Pick a question. Left: Gemma 3 1B before training. Right: makandal v2. Look at the language, the length, and whether the answer makes sense.
KonparezonComparison
Twa modèl, menm egzamen
Three models, the same tests
- Gemma 3 1Bpwen depa a, Googlethe starting point, by Google
- makandal v1aprann reponn yon sèl kesyonlearned to answer one question
- makandal v2konvèsasyon, zouti, matematik, syans, sante, jaden, lajan, lalwaconversations, tools, math, science, health, farming, money, law
Lekti, Belebele, 200 kesyon pa langReading, Belebele, 200 questions per language
| EgzamenTest | Gemma 3 1B | v1 | v2 |
|---|---|---|---|
| Chwazi bon zouti a, 300 demannPicks the right tool, 300 requests | 10% | – | 93% |
| Rele yon zouti lè sa nesesèCalls a tool when one is needed | 4% | – | 98% |
| Pa rele zouti pou ti kozeNo tool for small talk | 93% | – | 100% |
| Pwoblèm matematik an kreyòlKreyòl word problems | 6% | – | 30% |
| Repete tèt li nan yon konvèsasyonRepeats itself in a conversation | 6.6% | tonbe nan bouklooped | 0% |
| Erè sou repons kreyòl (pi ba, pi bon)Error on Kreyòl answers (lower is better) | 5.44 | 1.10 | 0.86 |
LimitLimits
Sa l poko konn fè
What it cannot do yet
- Fè san zoutiFacts without a tool"Ki kapital Ayiti?" te bay: "Ayiti gen anpil lajan. Kapital la se 1000000…""Ki kapital Ayiti?" got: "Ayiti gen anpil lajan. Kapital la se 1000000…"
- Zouti apre premye kesyon anTools after the first questionpi lwen nan konvèsasyon an, li konn reponn poukont li; v2.1 ap korije salater in a conversation it often answers on its own; v2.1 will fix this
- Pa pou konsèy gravNot for serious advicepa sèvi ak li poukont li pou sante, lalwa oswa lajando not rely on it alone for health, law or money
Yon modèl 1 milya paramèt pa ka kenbe tout konesans mond lan. Li bon pou konprann, pale, ak chèche; li pa bon pou sonje non, dat ak kote poukont li. Se pou sa zouti yo la.
A 1-billion-parameter model cannot hold all the world's knowledge. It is good at understanding, talking and looking things up; it is weak at remembering names, dates and places on its own. That is what the tools are for.
Eseye lTry it
Nan navigatè a, oswa sou machin ou
In the browser, or on your machine
Hugging Face Space
Pale ak Makandal ak uit zouti li yo: tan an, lè a, ansiklopedi, kalkil, memwa ak konesans pa w. Oswa poze menm kesyon an bay makandal v2 ak Gemma 3 1B, youn bò kote lòt.Talk to Makandal with its eight tools: weather, time, encyclopedia, arithmetic, memory and your own knowledge. Or ask makandal v2 and Gemma 3 1B the same question, side by side.
Louvri Space laOpen the Spacellama.cpp · GGUF
Q8_0 (1.07 GB, ~91 mo/s) oswa Q4_K_M (0.81 GB). --jinja limen zouti yo.Q8_0 (1.07 GB, ~91 tokens/s) or Q4_K_M (0.81 GB). --jinja turns the tools on.
llama-server -m makandal-multiple-v2-Q8_0.gguf -c 8192 --jinja
Python · transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "jsbeaudry/makandal-multiple-v2"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)
msgs = [{"role": "user", "content": "Kijan pou m plante pwa?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True,
return_tensors="pt", return_dict=True)
out = model.generate(**ids, max_new_tokens=120)
print(tok.decode(out[0, ids["input_ids"].shape[1]:],
skip_special_tokens=True))Kijan li fètHow it was made
Yon gwo modèl anseye yon ti modèl
A large model taught a small one
- 64,932egzanp, mwatye an kreyòlexamples, half in Kreyòl
- 107 minfòmasyon sou yon sèl GPUof training on one GPU
Yon modèl pwofesè, Gemma 4 26B, ekri repons nan stil yon asistan vwa, epi Gemma 4 31B tcheke yo. Sa ki pa kòrèk, ki pa nan bon lang, oswa ki pwomèt sa l pa ka fè, retire. 2.7% egzanp yo te ekri pa DeepSeek V4 Pro.
A teacher model, Gemma 4 26B, wrote answers in the style of a voice assistant, and Gemma 4 31B checked them. Anything wrong, in the wrong language, or promising what it cannot do was removed. 2.7% of the examples were written by DeepSeek V4 Pro.