AfriqueQwen3.5-9B-50Langs-Instruct-v1

Post-trained version (v1) of McGill-NLP/AfriqueQwen3.5-9B-50Langs.

Performance

Model Params AfriMGSM AfriMMLU AfriXNLI Belebele Injongo SIB-200 FLORES Overall
Apertus-8B-Instruct-2509 8B 12.7 35.1 34.4 42.5 11.8 67.7 43.5 35.4
Gemma-3-4B-it 4B 20.4 35.7 48.4 47.9 28.7 67.1 33.5 40.3
tiny-aya-earth 3.35B 19.6 41.3 55.4 54.4 17.4 69.5 42.0 42.8
Qwen3.5-4B 4B 41.5 41.8 51.5 52.8 35.1 71.5 44.4 48.4
Sunflower-Qwen3.5-9B 9B 19.9 41.8 50.5 62.7 55.0 78.3 62.3 52.9
AfriqueQwen3.5-4B-Instruct-v1 4B 51.4 46.3 62.4 65.2 59.1 80.4 57.8 60.4
AfriqueQwen3.5-4B-Instruct-v2 4B 50.8 46.4 63.6 66.2 58.7 81.2 62.3 61.3
AfriqueQwen3.5-4B-50Langs-Instruct-v1 4B 54.2 45.4 62.2 66.9 58.1 81.5 62.9 61.6
AfriqueQwen3.5-9B-50Langs-Instruct-v1 9B 64.6 51.2 67.8 73.2 70.8 86.4 63.8 68.3

For Flores, we use SSA-COMET (Li et al., 2025); for AfriMGSM, we detect the number inside the \boxed{} or the last number in the entire output; for multiple-choice questions, we use exact match.

Usage

Served with vLLM. The checkpoint declares Qwen3_5ForConditionalGeneration but carries no vision weights (426 language-model tensors plus an untied lm_head, 0 visual), so --language-model-only is required.

vllm serve McGill-NLP/AfriqueQwen3.5-9B-50Langs-Instruct-v1 --language-model-only --max-model-len 12288 --port 8000
curl http://127.0.0.1:8000/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "McGill-NLP/AfriqueQwen3.5-9B-50Langs-Instruct-v1",
  "messages": [
    {"role": "system", "content": "Wewe ni msaidizi mwenye manufaa ambaye hutatua matatizo ya hisabati yaliyoandikwa kwa maneno.\nTafadhali toa hoja hatua kwa hatua kwa Kiswahili, na uweke jibu lako la mwisho ndani ya \\boxed{}."},
    {"role": "user", "content": "Sandy hupata $15 kwa saa. Alifanya kazi saa 10 Ijumaa. Alipata pesa ngapi?"}
  ],
  "temperature": 0.7, "top_p": 0.8, "top_k": 20, "presence_penalty": 1.5, "max_tokens": 10240
}'
Hatua ya 1: Sandy hupata $15 kwa saa.
Hatua ya 2: Sandy alifanya kazi saa 10 Ijumaa.
Hatua ya 3: Jumla ya Mapato = $15/saa x saa 10
Hatua ya 4: $15 x 10 = $150

Sandy alipata $150 kwa kufanya kazi saa 10.

Recommendation Prompting and Setting

Mathematics

English

You are a helpful assistant that solves math word problems.
Please reason step by step in English, and put your final answer within \boxed{}.

For other languages, replace the part after "step by step in" with the target language, and translate it into the target language, for example Swahili:

Wewe ni msaidizi mwenye manufaa ambaye hutatua matatizo ya hisabati yaliyoandikwa kwa maneno.
Tafadhali toa hoja hatua kwa hatua kwa Kiswahili, na uweke jibu lako la mwisho ndani ya \boxed{}.
Setting Mathematics Other tasks
enable_thinking true false
temperature 0.7 0 (greedy)
top_p 0.8 —
top_k 20 —
min_p 0 —
presence_penalty 1.5 0
repetition_penalty 1.0 1.0
max_tokens 10240 1024
--max-model-len 12288 8192

Other tasks

Translation:

Translate the following sentence from {source language} to {target language}.

{source language}: {sentence}
{target language}:

Multiple choice:

Answer the multiple-choice question by replying with ONLY the letter
(A, B, C, or D) of the correct answer.

Question: {question}

A. {A}
B. {B}
C. {C}
D. {D}

Answer:
Downloads last month
2,202
Safetensors
Model size
9B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for McGill-NLP/AfriqueQwen3.5-9B-50Langs-Instruct-v1

Finetuned
(1)
this model
Finetunes
1 model
Quantizations
2 models

Space using McGill-NLP/AfriqueQwen3.5-9B-50Langs-Instruct-v1 1

Collection including McGill-NLP/AfriqueQwen3.5-9B-50Langs-Instruct-v1

Paper for McGill-NLP/AfriqueQwen3.5-9B-50Langs-Instruct-v1