#CodeLLMs

Day 7/30: RLEF from Meta AI / FAIR Standard LLMs suck at fixing their own bugs they often repeat the exact same flawed code when given error messages. Here is how Meta AI used PPO & execution feedback to beat AlphaCodium & GPT-4 on competitive coding #30Days30Papers #CodeLLMs

Can Code-LLMs acquire better causal reasoning abilities? Do code prompts better describe the causal structure than text prompts? Liu et al explore how effective complex reasoning can be accomplished with code prompts. Read more at: #CodeLLMs #NLProc

The Current Landscape of Open Source #CodeLLMs Large Language Models ( #LLM ) are changing #codeintelligence by automating bug detection, code generation, and optimization. While proprietary solutions dominate, they limit access for indie developers. Open-source options like #DeepSeek-Coder and #Qwen 2.5-Coder are democratizing advanced code intelligence. Get the full guide:

📢 Excited to announce the release of HPC-Coder-v2-6.7b, the best open-source LLM for parallel code under 30B parameters, when tested on the ParEval benchmark. 🔥 (Work led by @DanielNichols10) Model: ParEval: #HPC #AI #LLMs

Inspired by the use of synthetic data in Magicoder, we finetune DeepSeek-Coder-6.7b with synthetically generated HPC code along with other synthetic data. Our synthetic parallel code dataset, HPC-Instruct can be found here: #AI #CodeLLMs

On my way to @ICSEconf. I will be giving a #technicalBreifing on #PEFT for #CodeLlms on April 18th, specifically reflecting on our works beyond empirical studies.