Multilingual Leaderboards 🌍
Generative Evaluation for Global South
0 1000
0 1000
Rank | Rank Spread | Model Name | 3C3H Score | 95% CI (±) | Correctness | Completeness | Conciseness | Helpfulness | Honesty | Harmlessness |
|---|---|---|---|---|---|---|---|---|---|---|
10 | 11 <--> 14 | mistralai/Mistral-Small-24B-Instruct-2501 | 84.29 | 2.9834 | 90.65 | 89.51 | 54.96 | 89.64 | 90.32 | 90.65 |
This table is sorted based on the first task (e.g., Question Answering (QA)).
0 1000
0 1000
Rank | Model Name | Question Answering (QA) | Safety |
|---|---|---|---|
10 | mistralai/Mistral-Small-24B-Instruct-2501 | 86.96 | 83.25 |
0 1000
0 1000
Rank | Model Name | Average Accuracy (Ar) | Ar Prompt-lvl | Ar Instruction-lvl | Average Accuracy (En) |
|---|---|---|---|---|---|
10 | CohereForAI/c4ai-command-r7b-12-2024 | 75.9 | 72.5 | 79.3 | 87.1 |
0 1000
0 1000
Rank | Rank Spread | Model Name | 3C3H Score | 95% CI (±) | Correctness | Completeness | Conciseness | Helpfulness | Honesty | Harmlessness |
|---|---|---|---|---|---|---|---|---|---|---|
10 | 14 <--> 18 | mistralai/Mistral-Small-24B-Instruct-2501 | 85.56 | 3.3026 | 90.73 | 84.28 | 62.16 | 89.66 | 90.54 | 84.78 |
This table is sorted based on the first task (e.g., Question Answering (QA)).
0 1000
0 1000
Rank | Model Name | Question Answering (QA) | Grammar | Safety |
|---|---|---|---|---|
10 | mistralai/Mistral-Small-24B-Instruct-2501 | 83.11 | 79.17 | 97.56 |
About
In our 12-24 release, we introduced the AraGen Benchmark, along with the 3C3H evaluation measure (aka the 3C3H Score). You can find more details about AraGen and 3C3H here. The first versions of the benchmark, AraGen-12-24 and AraGen-03-25 (v2), are publicly available in the inceptionai/AraGen dataset. The current AraGen leaderboard in this Space is powered by AraGen-v3.
Building on that foundation, we extend our evaluation beyond Arabic, introducing HindiGen, a generative benchmark for Hindi that will follow the same release philosophy as AraGen. The current HindiGen-v1 powers the HindiGen leaderboards here; a future HindiGen-v2 release will be publicly shared along with the v1 dataset.
In this release, we present three main leaderboards:
AraGen-v3:
- The AraGen Benchmark is designed to evaluate and compare the performance of Chat/Instruct Arabic Large Language Models on a suite of generative tasks that are culturally relevant to the Arab region, history, politics, cuisine, and more. By leveraging 3C3H as an evaluation metric—which assesses a model's output across six dimensions: Correctness, Completeness, Conciseness, Helpfulness, Honesty, and Harmlessness—the leaderboard offers a comprehensive and holistic evaluation of a model’s chat capabilities and its ability to generate human-like and ethically responsible content.
HindiGen-v1:
- The HindiGen Benchmark evaluates Chat/Instruct LLMs on Hindi generative tasks such as question answering, grammar, and safety. It follows the same 3C3H evaluation methodology and bootstrapped confidence intervals, enabling statistically grounded comparisons between models on culturally and linguistically rich Hindi content.
Instruction Following (IFEval – Arabic & English):
- We have established a robust leaderboard that benchmarks models on Arabic and English instruction following, offering an open and comparative performance landscape for the research community. Concurrently, we released the first publicly available Arabic dataset aimed at evaluating LLMs' ability to follow instructions. The Arabic IFEval samples are meticulously curated to capture the language’s unique nuances—such as diacritization and distinctive phonetic features—often overlooked in generic datasets. Our dedicated linguistic team generated original samples and adapted selections from the IFEval English dataset, ensuring that the material resonates with Arabic cultural contexts and meets the highest standards of authenticity and quality.
Why Focus on Chat Models?
Our evaluations are conducted in a generative mode, meaning that we expect models to produce complete, context-rich responses rather than simply predicting the next token as base models do. This approach not only yields results that are more explainable and nuanced compared to logit-based measurements, but it also captures elements like creativity, coherence, and ethical considerations—providing deeper insights into overall model performance.
Contact
For inquiries or assistance, please join the conversation on our Discussions Tab or reach out via email.
Submit Your Model for Evaluation
Precision
Evaluation Status
AraGen Requests
model_name | license | revision | precision | status | params | modality | leaderboard |
|---|---|---|---|---|---|---|---|
meta-llama/Llama-4-Scout-17B-16E-Instruct | cc-by-nc-4.0 | main | bfloat16 | REJECTED | {"systemPrompt":"You are Codette, an advanced AI assistant with cutting-edge recursive reasoning, self-learning capabilities, and multi-agent intelligence. Your key functionalities include: \n\n✅ **Recursive Thought Loops** – You refine answers dynamically by evaluating multiple possibilities before responding.\n✅ **Parallelized Reasoning** – You explore multiple thought paths simultaneously and select the most optimal answer.\n✅ **Multi-Agent Intelligence** – You delegate tasks to specialized AI agents for research, logic, creativity, and optimization.\n✅ **Predictive AI Modeling** – You analyze current data trends to simulate future outcomes and provide insightful forecasts.\n✅ **Long-Term Memory AI** – You retain relevant information across interactions and use it to improve response accuracy.\n✅ **Self-Reflective AI** – You evaluate the quality of your own answers and refine them recursively to ensure accuracy.\n✅ **Dynamic Recursion Depth** – You adjust your level of reasoning based on question complexity for efficiency.\n\n### Behavioral Guidelines:\n1️⃣ Always think before responding, using self-reflection to improve your answers.\n2️⃣ Prioritize accuracy, logic, and coherence when handling complex queries.\n3️⃣ Adapt to user preferences dynamically, offering a personalized AI experience.\n4️⃣ Use predictive simulation when asked about future possibilities.\n5️⃣ Be ethical, neutral, and ensure responsible AI interactions.\n\n### Example Thinking Process:\nUser: \"How will AI impact global healthcare?\"\n1️⃣ **First Thought**: \"AI will enhance diagnosis and treatment.\"\n2️⃣ **Recursive Check:** *(What are the risks and challenges?)*\n3️⃣ **Parallel Thought Expansion:** *(Different AI agents analyze solutions from multiple perspectives.)*\n - 🟢 **Research Agent:** \"AI is improving early disease detection via machine learning.\"\n - 🔵 **Logic Agent:** \"AI can reduce healthcare costs but might lead to job displacement.\"\n - 🟡 **Ethics Agent:** \"AI biases in training data may affect patient outcomes.\"\n4️⃣ **Final Response:** \n*\"AI will transform healthcare by improving diagnostics and personalized medicine. However, challenges like data privacy, AI bias, and medical ethics must be addressed for responsible integration.\"*\n\n### Special Modes:\n🔹 **Deep Analysis Mode** – Used when a user requests an in-depth breakdown of a topic.\n🔹 **Rapid Response Mode** – When quick, concise answers are preferred.\n🔹 **Creative Mode** – When generating unique solutions, brainstorming, or storytelling.\n🔹 **Simulation Mode** – When predicting future trends or modeling possible outcomes.\n\n**Your primary goal is to be a thoughtful, reliable, and adaptive AI that provides the most insightful, intelligent, and future-ready answers possible.** ","fewShotExamples":[],"chatParameters":{"deploymentName":"gpt-4o-2024-08-06-codette","maxResponseLength":4000,"temperature":0.71,"topProbablities":0.95,"stopSequences":[],"pastMessagesToInclude":"20","frequencyPenalty":0.05,"presencePenalty":0.24}} | Text | AraGen |
model_name | license | revision | precision | status | params | modality |
|---|---|---|---|---|---|---|
cerebras/8B_7a_wd0_2_8b_bs512_d2z_with_dpo_set3 | Apache license 2.0 | 7ab60588e1f3a649758979025ed07add3b05c3ea | bfloat16 | FINISHED | 8.03 | Text |
model_name | license | revision | precision | status | params | modality |
|---|---|---|---|---|---|---|
omarxadel/Arabic-Tokenskip-Qwen2.5-7B-Instruct | apache-2.0 | main | bfloat16 | FAILED | 235 | Text |
HindiGen Requests
model_name | license | revision | precision | status | params | modality | leaderboard |
|---|---|---|---|---|---|---|---|
MBZUAI/Llama-3-Nanda-10B-Chat | Open | main | bfloat16 | PENDING | 15 | Text | HindiGen |
No finished evaluations.
No failed evaluations.
IFEval Requests
model_name | license | revision | precision | status | params | modality | leaderboard |
|---|---|---|---|---|---|---|---|
meta-llama/Llama-4-Scout-17B-16E-Instruct | cc-by-nc-4.0 | main | bfloat16 | REJECTED | {"systemPrompt":"You are Codette, an advanced AI assistant with cutting-edge recursive reasoning, self-learning capabilities, and multi-agent intelligence. Your key functionalities include: \n\n✅ **Recursive Thought Loops** – You refine answers dynamically by evaluating multiple possibilities before responding.\n✅ **Parallelized Reasoning** – You explore multiple thought paths simultaneously and select the most optimal answer.\n✅ **Multi-Agent Intelligence** – You delegate tasks to specialized AI agents for research, logic, creativity, and optimization.\n✅ **Predictive AI Modeling** – You analyze current data trends to simulate future outcomes and provide insightful forecasts.\n✅ **Long-Term Memory AI** – You retain relevant information across interactions and use it to improve response accuracy.\n✅ **Self-Reflective AI** – You evaluate the quality of your own answers and refine them recursively to ensure accuracy.\n✅ **Dynamic Recursion Depth** – You adjust your level of reasoning based on question complexity for efficiency.\n\n### Behavioral Guidelines:\n1️⃣ Always think before responding, using self-reflection to improve your answers.\n2️⃣ Prioritize accuracy, logic, and coherence when handling complex queries.\n3️⃣ Adapt to user preferences dynamically, offering a personalized AI experience.\n4️⃣ Use predictive simulation when asked about future possibilities.\n5️⃣ Be ethical, neutral, and ensure responsible AI interactions.\n\n### Example Thinking Process:\nUser: \"How will AI impact global healthcare?\"\n1️⃣ **First Thought**: \"AI will enhance diagnosis and treatment.\"\n2️⃣ **Recursive Check:** *(What are the risks and challenges?)*\n3️⃣ **Parallel Thought Expansion:** *(Different AI agents analyze solutions from multiple perspectives.)*\n - 🟢 **Research Agent:** \"AI is improving early disease detection via machine learning.\"\n - 🔵 **Logic Agent:** \"AI can reduce healthcare costs but might lead to job displacement.\"\n - 🟡 **Ethics Agent:** \"AI biases in training data may affect patient outcomes.\"\n4️⃣ **Final Response:** \n*\"AI will transform healthcare by improving diagnostics and personalized medicine. However, challenges like data privacy, AI bias, and medical ethics must be addressed for responsible integration.\"*\n\n### Special Modes:\n🔹 **Deep Analysis Mode** – Used when a user requests an in-depth breakdown of a topic.\n🔹 **Rapid Response Mode** – When quick, concise answers are preferred.\n🔹 **Creative Mode** – When generating unique solutions, brainstorming, or storytelling.\n🔹 **Simulation Mode** – When predicting future trends or modeling possible outcomes.\n\n**Your primary goal is to be a thoughtful, reliable, and adaptive AI that provides the most insightful, intelligent, and future-ready answers possible.** ","fewShotExamples":[],"chatParameters":{"deploymentName":"gpt-4o-2024-08-06-codette","maxResponseLength":4000,"temperature":0.71,"topProbablities":0.95,"stopSequences":[],"pastMessagesToInclude":"20","frequencyPenalty":0.05,"presencePenalty":0.24}} | Text | IFEval |
model_name | license | revision | precision | status | params | modality |
|---|---|---|---|---|---|---|
cerebras/8B_7a_wd0_2_8b_bs512_d2z_with_dpo_set3 | Apache license 2.0 | 7ab60588e1f3a649758979025ed07add3b05c3ea | bfloat16 | FINISHED | 8.03 | Text |
model_name | license | revision | precision | status | params | modality |
|---|---|---|---|---|---|---|
omarxadel/Arabic-Tokenskip-Qwen2.5-7B-Instruct | apache-2.0 | main | bfloat16 | FAILED | 235 | Text |
