Introducing: StarCoder2 and The Stack v2 ⭐️
StarCoder2 is trained with a 16k token context and repo-level information for 4T+ tokens. All built on The Stack v2 - the largest code dataset with 900B+ tokens.
All code, data and models are fully open!
hf.co/bigcode/starcoder2-15b
Introducing BigCodeArena, a human-in-the-loop platform for evaluating code through execution.
Unlike current open evaluation platforms that collect human preferences on text, it enables interaction with runnable code to assess functionality and quality across any language.
BigCodeArena cannot be built without the support of the BigCode community. We are grateful for the huge credits provided by the @e2b team. We thank @hyperbolic_labs, @nvidia, and @Alibaba_Qwen for providing the model inference endpoints.
Beyond human votes, we release two new benchmarks:
- BigCodeReward: tests reward models on 4.7K human preference votes. Execution feedback improves judgment accuracy.
- AutoCodeArena: automated evaluation of 20+ LLMs. GPT-5 leads, followed by Claude-Opus-4 and Claude-Sonnet-4.
In 5 months, BigCodeArena collected 14K conversations & 4.7K preference votes across 10 frontier LLMs.
Findings:
1. o3-mini & o1-mini consistently top Elo rankings
2. Claude-3.5-Sonnet excels in matched-language settings
3. Previous open models still lag
Why does this matter?
Benchmarks like HumanEval only scratch the surface. Reading code “by eye” is error-prone. True quality emerges when you actually run it: web apps render, games play, edge cases break.
BigCodeArena makes execution feedback the default.
Introducing BigCodeArena, a human-in-the-loop platform for evaluating code through execution.
Unlike current open evaluation platforms that collect human preferences on text, it enables interaction with runnable code to assess functionality and quality across any language.
BigCodeBench @BigCodeProject evaluation framework has been fully upgraded! Just pip install -U bigcodebench
With v0.2.0, it's now much easier to use compared to the previous v0.1.* versions. The new version adopts the @Gradio Client API interface from @huggingface Spaces by default, w/o the need for local environment setup, and can be replaced with a custom API if desired. Moreover, the latest version no longer requires running separate commands for each stage (like generate, sanitize, and evaluate), significantly simplifying the workflow.
The new version also features Batch Inference — running the LLMs on the BigCodeBench-Full set now takes under 5mins for generation and execution!
BTW, the benchmark data has been updated to v0.1.2, improving task instructions and test examples.
Some of the updates in this release were inspired by EvalPlus @JiaweiLiu_ . A big thank you for the continued maintenance of EvalPlus and the strong support for BigCodeBench 🤗
Evaluating LM agents has come a long way since gpt-4 released in March of 2023.
We now have SWE-Bench, (Visual) Web Arena, and other evaluations that tell us a lot about how the best models + architectures do on hard and important tasks.
There's still lots to do, though 🧵
People may think BigCodeBench @BigCodeProject is nothing more than a straightforward coding benchmark, but it is not.
BigCodeBench is a rigorous testbed for LLM agents using code to solve complex and practical challenges. Each task demands significant reasoning capabilities for selecting appropriate library APIs and logically connecting them to craft a program.
Rather than merely providing high-level instructions, each task comes with detailed requirements to evaluate the model's ability to adhere to all aspects. While language models typically perform well on short and simple tasks, they often struggle with longer and more complex problems (e.g., BigCodeBench-Hard).
A model needs to perform well on BigCodeBench before it can be used for agentic software development.
LLMs are great at programming tasks... for Python and other very popular PLs. But, they are often unimpressive at artisanal PLs, like OCaml or Racket. We've come up with a way to significantly boost LLM performance of on low-resource languages. If you care about them, read on!
Today, we are happy to announce the beta mode of real-time Code Execution for BigCodeBench @BigCodeProject, which has been integrated into our Hugging Face leaderboard.
We understand that setting up a dependency-based execution environment can be cumbersome, even with the built-in Docker image and Dockerfile. To make the evaluation process more reproducible, we've built an interactive environment for you, with guidance from the @Gradio team! (Special thanks to @evilpingwin 🤗)
Please note: (1) The execution process might be slightly slower than what you experience on a local machine, as we are using the basic CPU option. There are some compatibility issues with the upgraded CPU environment, and we are currently exploring solutions. (2) Four tasks in the full set require some tricky setup, which has resulted in a pass rate of 99.6%. We will work to fix these in the next iterations :)
In the past few months, we’ve seen SOTA LLMs saturating basic coding benchmarks with short and simplified coding tasks. It's time to enter the next stage of coding challenge under comprehensive and realistic scenarios!
-- Here comes BigCodeBench, benchmarking LLMs on solving practical and challenging programming tasks!
So, can LLMs solve these tasks?
- Not yet!
🏆 Pass@1: Humans ace 97%, GPT-4o only hits 50-60%, but DeepSeek-Coder-V2 is tighy at its heels!
Check out our leaderboard, data, code, and paper: bigcode-bench.github.io/
1/🧵
Releasing BigCodeBench-Hard: a subset of more challenging and user-facing tasks.
BigCodeBench-Hard provides more accurate model performance evaluations and we also investigate some recent model updates.
Read more: huggingface.co/blog/terryyz/…
Leaderboard: huggingface.co/spaces/bigcod…