StarCoder 2 Code Generation LLM (BigCode 2024年)（スターコーダー2）
2024年BigCode (Hugging Face+ServiceNow)発表StarCoder 2・Industry-leading open-source code generation LLM + Industry-leading 3B/7B/15B + Industry-leading 619 programming languages + Industry-leading The Stack v2 dataset。
概要
StarCoder 2 Code Generation LLM は、2024年BigCode (Hugging Face+ServiceNow)発表StarCoder 2・Industry-leading open-source code generation LLM 2024年 + Industry-leading The Stack v2 dataset position確立。StarCoder 2 specifications = Industry-leading open-source code generation LLM (Industry-leading StarCoder 2 2024 + Industry-leading open-source code generation + Industry-leading StarCoder 2 BigCode flagship) + Industry-leading 3B/7B/15B variants (Industry-leading 3B + 7B + 15B 3-tier sizes) + Industry-leading 619 programming languages (Industry-leading 619 programming languages support) + Industry-leading The Stack v2 dataset (Industry-leading The Stack v2 67TB dataset)。
主な特徴・仕組み
- Org: Industry-leading BigCode (Hugging Face + ServiceNow + NVIDIA collaboration)
- Year: 2024年 (arXiv 2024年2月発表)
- Paper: Industry-leading "StarCoder 2 and The Stack v2: The Next Generation" arXiv 2402.19173
- Sizes: Industry-leading 3B + 7B + 15B 3-tier variants
- Languages: Industry-leading 619 programming languages support
- Dataset: Industry-leading The Stack v2 67TB code dataset
- Context Length: Industry-leading 16K tokens context
- License: Industry-leading BigCode Open RAIL-M license (commercial use OK)
- Training Tokens: Industry-leading 4T+ tokens training
- Open Source: Industry-leading StarCoder 2 HuggingFace bigcode/starcoder2-15b
- Fill-in-the-Middle: Industry-leading FIM training enabled
- Industry-Leading: Industry-leading open-source code generation LLM BigCode 2024
スペック比較表
| LLM Code Generation (2023-2024) | Year | Org | Sizes | Industry Position |
|---|---|---|---|---|
| StarCoder 2 | 2024 | BigCode HF+ServiceNow | 3B/7B/15B 619 langs | Industry-leading BigCode open |
| CodeLlama | 2023 | Meta AI | 7B/13B/34B/70B | Industry-leading Meta CodeLlama |
| DeepSeek-Coder | 2023 | DeepSeek AI | 1.3B-33B + V2 236B MoE | Industry-leading DeepSeek-Coder V2 |
| Qwen2.5-Coder | 2024 | Alibaba Qwen | 0.5B/1.5B/3B/7B/14B/32B | Industry-leading Qwen2.5-Coder |
| Codestral | 2024 | Mistral AI | 22B / Mamba 7B | Industry-leading Mistral Codestral |
具体例・対応製品
- StarCoder 2 (2024年, BigCode HF+ServiceNow): Industry-leading open-source code generation
- Industry-leading 3B + 7B + 15B 3-tier variants: Industry-leading 3B/7B/15B
- Industry-leading 619 programming languages: Industry-leading 619 languages
- Industry-leading The Stack v2 67TB code dataset: Industry-leading The Stack v2 67TB
- Industry-leading BigCode Open RAIL-M license commercial use OK: Industry-leading Open RAIL-M
- 競合 CodeLlama + DeepSeek-Coder + Qwen2.5-Coder + Codestral: Industry-leading code gen competitors
自作PCでの選び方・注意点
StarCoder 2 Code Generation LLM は「Industry-leading open-source code generation LLM + Industry-leading 3B/7B/15B」「Industry-leading 619 programming languages + Industry-leading The Stack v2 dataset」用途のIndustry-leading open-source code generation 2024年BigCode発表product。Industry-leading 3-tier (Industry-leading 3B + 7B + 15B 3-tier variants + Industry-leading StarCoder 2 3-tier signature) で Industry-leading 3-tier + Industry-leading StarCoder 2 signature。Industry-leading 619 languages (Industry-leading 619 programming languages support + Industry-leading StarCoder 2 multilang signature + Industry-leading widest language support) で Industry-leading 619 languages + Industry-leading multilang signature + Industry-leading widest support。Industry-leading The Stack v2 (Industry-leading The Stack v2 67TB code dataset + Industry-leading BigCode dataset signature + Industry-leading 4T+ tokens training) で Industry-leading The Stack v2 67TB + Industry-leading BigCode dataset + Industry-leading 4T+ tokens。Industry-leading BigCode Open RAIL-M (Industry-leading BigCode Open RAIL-M license commercial use OK + Industry-leading FIM Fill-in-the-Middle training + Industry-leading 16K context) で Industry-leading Open RAIL-M + Industry-leading FIM training + Industry-leading 16K context。Industry-leading BigCode collaboration (Industry-leading BigCode Hugging Face + ServiceNow + NVIDIA collaboration + Industry-leading StarCoder 2 HuggingFace bigcode/starcoder2-15b + Industry-leading open-source code generation LLM BigCode 2024) で Industry-leading BigCode HF+ServiceNow+NVIDIA + Industry-leading bigcode/starcoder2-15b + Industry-leading BigCode 2024。但しIndustry-leading CodeLlama + DeepSeek-Coder + Qwen2.5-Coder + Codestral competition (Industry-leading CodeLlama Meta 2023 + DeepSeek-Coder V2 2023 + Qwen2.5-Coder Alibaba 2024 + Codestral Mistral 2024 vs StarCoder 2 BigCode 619 langs The Stack v2 Open RAIL-M 2024 trade-off) で Industry-leading 4-code generation model competitors vs StarCoder 2 + Industry-leading 3B/7B/15B + 619 programming languages + The Stack v2 67TB dataset + FIM training + 16K context + BigCode Open RAIL-M license + BigCode HF+ServiceNow+NVIDIA collaboration unique advantage adoption alignment必須。
関連用語との違い
- vs CodeLlama (Meta 2023): StarCoder 2はBigCode + 3B/7B/15B + 619 langs + Open RAIL-M・CodeLlamaはMeta + 7B/13B/34B/70B + Llama Community License
- vs DeepSeek-Coder (DeepSeek 2023): StarCoder 2は619 langs + Stack v2・DeepSeek-CoderはV2 MoE 236B + 86 langs
- vs Codestral (Mistral 2024): StarCoder 2はBigCode open-source・CodestralはMistral Non-Production License
よくある質問（FAQ）
Q1: StarCoder 2 vs CodeLlama 違いは？ A: StarCoder 2 (Industry-leading 3B + 7B + 15B 3-tier variants + 619 programming languages + The Stack v2 67TB code dataset + 16K tokens context + 4T+ tokens training + FIM Fill-in-the-Middle + BigCode Open RAIL-M license + BigCode Hugging Face + ServiceNow + NVIDIA 2024) vs CodeLlama (Industry-leading 7B + 13B + 34B + 70B + Llama 2 base + 16K-100K context Code Llama variants + Python-specialized + Instruct variants + Meta AI 2023 + Llama Community License)・Industry-leading BigCode + 3B/7B/15B + 619 langs + Stack v2 + Open RAIL-M = StarCoder 2 + Industry-leading Meta + 7B-70B + Llama 2 + Python + Instruct = CodeLlama preference judgment。
Q2: Industry-leading 619 programming languages + The Stack v2 value は？ A: Industry-leading 619 langs + Stack v2 (Industry-leading 619 programming languages support + Industry-leading widest language support + Industry-leading The Stack v2 67TB code dataset + Industry-leading 4T+ tokens training)。
Q3: Industry-leading BigCode Open RAIL-M + FIM value は？ A: Industry-leading Open RAIL-M + FIM (Industry-leading BigCode Open RAIL-M license commercial use OK + Industry-leading FIM Fill-in-the-Middle training enabled + Industry-leading 16K tokens context + Industry-leading BigCode HF+ServiceNow+NVIDIA collaboration)。
まとめ
StarCoder 2 = 2024年BigCode Hugging Face + ServiceNow + NVIDIA発表のopen-source code generation LLM。Industry-leading 3B + 7B + 15B 3-tier variants + Industry-leading 619 programming languages support + Industry-leading The Stack v2 67TB code dataset + Industry-leading 4T+ tokens training + Industry-leading 16K tokens context + Industry-leading FIM Fill-in-the-Middle training + Industry-leading BigCode Open RAIL-M license commercial use OK + Industry-leading BigCode Hugging Face + ServiceNow + NVIDIA collaboration + Industry-leading StarCoder 2 HuggingFace bigcode/starcoder2-15b + Industry-leading open-source code generation LLM BigCode 2024 position確立。