32 211 200

Aymeric Roucher

m-ric

http://aymeric-roucher.github.io

AI & ML interests

MLE at Hugging Face 🤗 LLMs, Agents, RAG, Multimodal.

Articles

Organizations

Posts 40

Post

951

🔥 𝐐𝐰𝐞𝐧 𝐫𝐞𝐥𝐞𝐚𝐬𝐞𝐬 𝐭𝐡𝐞𝐢𝐫 𝟐.𝟓 𝐟𝐚𝐦𝐢𝐥𝐲 𝐨𝐟 𝐦𝐨𝐝𝐞𝐥𝐬: 𝐍𝐞𝐰 𝐒𝐎𝐓𝐀 𝐟𝐨𝐫 𝐚𝐥𝐥 𝐬𝐢𝐳𝐞𝐬 𝐮𝐩 𝐭𝐨 𝟕𝟐𝐁!

The Chinese LLM maker just dropped a flurry of different models, ensuring there will be a Qwen SOTA model for every application out there:
Qwen2.5: 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B
Qwen2.5-Coder: 1.5B, 7B, and 32B on the way
Qwen2.5-Math: 1.5B, 7B, and 72B.

And they didn't sleep: the performance is top of the game for each weight category!

𝐊𝐞𝐲 𝐢𝐧𝐬𝐢𝐠𝐡𝐭𝐬:

🌐 All models have 𝟭𝟮𝟴𝗸 𝘁𝗼𝗸𝗲𝗻 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗹𝗲𝗻𝗴𝘁𝗵

📚 Models pre-trained on 18T tokens, even longer than the 15T of Llama-3

💪 The flagship 𝗤𝘄𝗲𝗻𝟮.𝟱-𝟳𝟮𝗕 𝗶𝘀 ~𝗰𝗼𝗺𝗽𝗲𝘁𝗶𝘁𝗶𝘃𝗲 𝘄𝗶𝘁𝗵 𝗟𝗹𝗮𝗺𝗮-𝟯.𝟭-𝟰𝟬𝟱𝗕, 𝗮𝗻𝗱 𝗵𝗮𝘀 𝗮 𝟯-𝟱% 𝗺𝗮𝗿𝗴𝗶𝗻 𝗼𝗻 𝗟𝗹𝗮𝗺𝗮-𝟯.𝟭-𝟳𝟬𝗕 𝗼𝗻 𝗺𝗼𝘀𝘁 𝗯𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸𝘀.

🇫🇷 On top of this, it 𝘁𝗮𝗸𝗲𝘀 𝘁𝗵𝗲 #𝟭 𝘀𝗽𝗼𝘁 𝗼𝗻 𝗺𝘂𝗹𝘁𝗶𝗹𝗶𝗻𝗴𝘂𝗮𝗹 𝘁𝗮𝘀𝗸𝘀 so it might become my standard for French

💻 Qwen2.5-Coder is only 7B but beats competing models up to 33B (DeeSeek-Coder 33B-Instruct). Let's wait for their 32B to come out!

🧮 Qwen2.5-Math sets a new high in the ratio of MATH benchmark score to # of parameters. They trained it by "aggregating more high-quality mathematical data, particularly in Chinese, from web sources, books, and codes across multiple recall cycles."

📄 Technical report to be released "very soon"

🔓 All models have the most permissive license apache2.0, except the 72B models that have a custom license mentioning "you can use it for free EXCEPT if your product has over 100M users"

🤗 All models are available on the HF Hub! ➡️ Qwen/qwen25-66e81a666513e518adb90d9e

Post

1497

𝗔𝗿𝗲 𝗔𝗴𝗲𝗻𝘁𝘀 𝗰𝗮𝗽𝗮𝗯𝗹𝗲 𝗲𝗻𝗼𝘂𝗴𝗵 𝗳𝗼𝗿 𝗗𝗮𝘁𝗮 𝗦𝗰𝗶𝗲𝗻𝗰𝗲? ⇒ 𝗠𝗲𝗮𝘀𝘂𝗿𝗲 𝘁𝗵𝗲𝗶𝗿 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝘄𝗶𝘁𝗵 𝗗𝗦𝗕𝗲𝗻𝗰𝗵 📊

A team from Tencent AI wanted to evaluate agentic systems on data science (DS) tasks : but they noticed that existing agentic benchmarks were severely limited in several aspects: they were limited to text and did not include tables or images, were only specific to certain packages, only performed exact match evaluation…

➡️ So they set out to build a much more exhaustive approach, to finally make the definitive DS agent benchmark.

𝗧𝗵𝗲 𝗗𝗦𝗕𝗲𝗻𝗰𝗵 𝗱𝗮𝘁𝗮𝘀𝗲𝘁
▪️DS bench has 466 data analysis tasks and 74 data modelling tasks
▪️The tasks are sourced from ModelOff and Kaggle, the platforms hosting the most popular data science competitions
▪️Difference with previous DS benchmarks:
❶ This benchmark leverages various modalities on top of text: images, Excel files, tables
❷ Complex tables: sometimes several tables should be leveraged to answer one question
❸ The context is richer, with longer descriptions.
▪️ Evaluation metrics : the benchmark is scored with an LLM as a judge, using a specific prompt.

𝗜𝗻𝘀𝗶𝗴𝗵𝘁𝘀 𝗳𝗿𝗼𝗺 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁𝘀
▪️ Their evaluation confirms that using LLMs in an agent setup, for instance by allowing them to run a single step of code execution, is more costly (especially with multi-turn frameworks like autogen) but also much more performant than the vanilla LLM.
▪️ The sets of tasks solved by different models (like GPT-3.5 vs Llama-3-8B) has quite low overlap, which suggests that different models tend to try very different approches.

This new benchmark is really welcome, can't wait to try transformers agents on it! 🤗

Read their full paper 👉 DSBench: How Far Are Data Science Agents to Becoming Data Science Experts? (2409.07703)

View all posts