# BigCode

> Source: https://aiwiki.ai/wiki/bigcode
> Updated: 2026-11-18
> Fact-checked: 2026-11-18
> Categories: AI Code Generation, Open Source AI, Research Organizations
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "BigCode." aiwiki.ai, 18 Nov 2026. https://aiwiki.ai/wiki/bigcode
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**BigCode** is an open scientific collaboration that works on the responsible development and use of [large language models](https://aiwiki.ai/wiki/large_language_model) for code (code LLMs), jointly led by [Hugging Face](https://aiwiki.ai/wiki/hugging_face) and [ServiceNow](https://aiwiki.ai/wiki/servicenow).[1][2] Announced on September 26, 2022,[4] the project produced The Stack, a large dataset of permissively licensed source code, and the code models [SantaCoder](https://aiwiki.ai/wiki/santacoder), [StarCoder](https://aiwiki.ai/wiki/starcoder) and StarCoder2,[3][13] along with evaluation tools such as the bigcode-evaluation-harness and the [BigCodeBench](https://aiwiki.ai/wiki/bigcodebench) benchmark.[19][21] The project also documented its governance in detail: it filtered training data by license, gave developers a tool to check whether their code had been collected and a process to have it removed, redacted personal information from training data, and released its models under OpenRAIL-M licenses (StarCoder and StarCoder2 under the BigCode OpenRAIL-M license), which allow commercial use but attach use restrictions.[3][9]

The project was inspired by BigScience, the open scientific collaboration that released the [BLOOM](https://aiwiki.ai/wiki/bloom) multilingual language model in July 2022.[3][5][8] The StarCoder paper described the community in 2023 as "co-stewarded by two industry research labs" with more than 600 members, and by the StarCoder2 release in February 2024 the project reported more than 1,200 members from more than 60 countries.[9][13]

## Origins

According to the BigCode governance card, a 2023 paper in which the project documents its own structure and decisions, the idea came out of a discussion in Utrecht that Harm de Vries of ServiceNow Research started with Thomas Wolf of Hugging Face. De Vries was inspired by the BigScience project and proposed a jointly led open collaboration on code models with the wider machine learning and open source communities.[3] A research collaboration agreement between ServiceNow and Hugging Face set the terms. Its scope covered collecting publicly accessible code and running license detectors on it, training dense transformer models with the [Megatron-LM](https://aiwiki.ai/wiki/megatron_lm) architecture, training retrieval-augmented code generation models, and building an evaluation suite.[3]

The original technical goal, set before the public announcement, was to train and release a 12-billion-parameter model that matched the [Codex](https://aiwiki.ai/wiki/openai_codex) model described by [OpenAI](https://aiwiki.ai/wiki/openai) in 2021, which OpenAI had not released; the card notes that it was offered through an API as code-cushman-001, though it was not entirely clear whether that model matched the one in the paper. The organizers planned to compare against Codex on [HumanEval](https://aiwiki.ai/wiki/humaneval) and APPS, but along the way concluded that code LLMs needed a broader evaluation suite.[3] Before the launch, a ServiceNow team spent several months inventorying active [GitHub](https://aiwiki.ai/wiki/github) repositories, downloading them, filtering out large files and duplicates, and detecting each repository's license. That work became The Stack.[3]

TechCrunch's report on the launch said the organizers would train a 15-billion-parameter model on ServiceNow's in-house GPU cluster, that only files from repositories with permissive licenses would go into the training data, and that the project was open to people with a professional AI research background who could commit time to it.[5] The article placed the launch against criticism of [GitHub Copilot](https://aiwiki.ai/wiki/github_copilot), noting that the Software Freedom Conservancy had objected to GitHub and OpenAI training Codex on public code that was not all permissively licensed.[5]

<figure class="embedvideo">
<iframe src="https://www.youtube-nocookie.com/embed/4qm7UPFMv64" title="BigCode: Open and responsible development of LLMs for code"></iframe>
<figcaption>"BigCode: Open and responsible development of LLMs for code", official video from ServiceNow, published June 18, 2024. <a href="https://www.youtube.com/watch?v=4qm7UPFMv64">Watch on YouTube</a>.</figcaption>
</figure>

## Mission and organization

BigCode's mission statement says that researchers working on code LLMs face "the lack of transparency around the development of these systems", which it says hinders outside accountability and keeps all but a few well-funded labs from taking a meaningful part in shaping the technology.[1] The project invited researchers to work on three areas in particular:[1]

- building a representative evaluation suite for code LLMs across many tasks and programming languages;
- developing faster methods for training and inference of LLMs;
- the legal, ethics, governance and safety aspects of code LLMs.

ServiceNow and Hugging Face committed research, engineering, ethics, governance and legal staff to the project. ServiceNow Research and Hugging Face made their compute clusters available for training, and Hugging Face hosts the project's datasets, models and demos on its Hub.[2] The mission page says that, while the companies provide this corporate support, "all technical governance takes place within working groups and task forces across the community".[1]

### Governance

A steering committee jointly led by ServiceNow and Hugging Face organizes the project, sets research strategy and publication goals, and oversees the working groups.[2] The governance card names Harm de Vries (ServiceNow) and Leandro von Werra (Hugging Face) as its leads, supported by a core team of Raymond Li, Denis Kocetkov and Sean Hughes from ServiceNow and Carlos Muñoz Ferrandis, Loubna Ben Allal and Thomas Wolf from Hugging Face; Yacine Jernite, Armel Randy and Joel Lamy-Poirier joined the core team later.[3] Decisions that cannot be settled at the community level go to the relevant working group lead, with the steering committee breaking ties as a last resort.[2] The SantaCoder report describes working groups on datasets, training methods, evaluation and ethics; its Legal, Ethics, and Governance working group took up data licensing, attribution of generated code, removal of personally identifiable information (PII) and the risk of models producing malicious code.[8]

### Membership

Participation is open to people who have a professional research background and can commit time to the project, generally people affiliated with an academic or industry research organization.[2] Members are required to follow the BigCode code of conduct and to understand the project's intellectual property rules.[3] The project coordinates work through a GitHub project board and Slack, trains models with a clone of Megatron-LM, keeps code repositories on GitHub and hosts weights and datasets on Hugging Face.[2]

| Point in time | Reported size | Source |
|---|---|---|
| May 4, 2023 | 675 participants, 629 of them members, from 62 countries | Governance card[3] |
| 2023 (StarCoder paper) | More than 600 members | StarCoder paper[9] |
| February 2024 | More than 1,200 members from more than 60 countries; 417 active between 2023-10-27 and 2024-02-24 | StarCoder2 paper[13] |

As of May 2023 the five largest national groups were the United States (222), India (60), the United Kingdom (36), Canada (35) and Germany (30), and the community used 48 Slack channels for the steering committee, working groups, task forces and general discussion.[3]

### Resources and costs

The governance card is unusually specific about what the project cost. ServiceNow collected a raw dataset of 5.28 billion files totaling 92 TB, from which The Stack was filtered. SantaCoder was trained on the ServiceNow cluster with 96 Tesla V100 GPUs, and StarCoder on a Hugging Face cluster with 512 A100 80 GB GPUs across 64 nodes. The card estimates SantaCoder's training emitted 124 kg of CO2-equivalent and StarCoderBase's 16.68 tonnes.[3] It puts the host companies' staff commitment at the equivalent of six full-time employees for the duration of the project and estimates that volunteer time at least matched it. ServiceNow Research spent about $6,000 on a community meetup at NeurIPS 2022, and Hugging Face paid $39,000 to crowd workers for data annotation through Toloka, which waived its platform fees.[3]

## Datasets

### The Stack

The first milestone was [The Stack](https://aiwiki.ai/wiki/the_stack), introduced on October 27, 2022, with a paper posted to arXiv in November 2022.[4][6] The first release contained 3.1 TB of permissively licensed source code in 30 programming languages. The paper reported that near-deduplication significantly improved results and that 350M-parameter models trained only on permissively licensed Python data could match previously reported HumanEval and MBPP scores.[6]

BigCode's dataset page records three version-one releases:[7]

| Release | Changes |
|---|---|
| v1.0 | 30 languages and 18 licenses, three of them (MPL, EPL, LGPL) weak copyleft; about 3 TB after near-deduplication |
| v1.1 | The three weak copyleft licenses removed; 193 permissive licenses and 358 languages; opt-out requests received by November 15, 2022 excluded; about 6 TB after near-deduplication |
| v1.2 | Opt-out requests received by February 9, 2023 excluded, along with files first flagged as malicious |

Alongside the code, BigCode released a collection of GitHub issues, repository metadata and other datasets derived from the project's research.[7] Anyone downloading The Stack must accept its terms of use, which require users to respect the original licenses of the code (including attribution clauses), to update to the latest version so that removals take effect, and to pass the same terms on to anyone they share the data with.[3] The governance card reports that The Stack had been downloaded 50,200 times as of May 3, 2023.[3]

### The Stack v2

[The Stack v2](https://aiwiki.ai/wiki/the_stack_v2), released with StarCoder2 on February 28, 2024, was built in partnership with Software Heritage on top of that organization's archive of source code, which covers 619 programming languages. BigCode added GitHub pull requests, [Kaggle](https://aiwiki.ai/wiki/kaggle) notebooks and code documentation, giving a training set four times larger than the one used for the first StarCoder.[13] Software Heritage was launched by [Inria](https://aiwiki.ai/wiki/inria) in partnership with UNESCO.[15] For transparency about provenance, the project released the Software Heritage persistent identifiers (SWHIDs) of the source files used.[13]

### The Stack v3

In 2026 Hugging Face researchers published The Stack v3 under a separate Hugging Face organization, HuggingFaceCode. Its dataset card calls it the successor to The Stack v2 and credits the BigCode community's work on dataset filters, PII detection and code dataset governance.[24] Unlike v2, which drew on the Software Heritage graph, v3 is a direct crawl of GitHub completed on August 7, 2025. Its training subset contains 15.9 TB of code in 713 languages from 173 million repositories, about 3.6 trillion tokens, and the file contents are included in the dataset itself. It redacts personal information with BigCode's StarPII model and sends removal requests through BigCode's opt-out repository.[24]

## Data governance

### Consent and opt-out

BigCode stated a goal of giving developers "agency over their source code" when it came to model training.[7] It started from license selection: The Stack kept code from repositories with open source licenses, and the governance card says the team kept only licenses without an attribution clause, because a model's output cannot reliably be attributed to the training data it came from.[3] The Stack's own terms nonetheless require anyone using the data to honour attribution clauses where the original licenses have them.[3][7] Because many of those licenses predate code LLMs, the project added an opt-out on top.[3]

The "Am I in The Stack?" tool, launched on November 15, 2022, lets developers search the dataset for their repositories and then file a removal request on GitHub. Requests are validated by checking that the GitHub username filing them matches the owner of the repositories, and validated requests are stored so that the code stays out of future versions.[3][4] At the time the StarCoder training data was processed, 44 people had opted out.[3] Before StarCoder2 training, BigCode updated the tool for The Stack v2 and set a deadline of November 20, 2023; it then removed 1,561 repositories belonging to 91 users and organizations, 22,066 files in all.[13] The opt-out repository on GitHub was still receiving requests in November 2026, now covering The Stack v3 as well.[25]

The project also gathered feedback on the approach. Between January and March 2023 members ran interviews and co-design workshops with 50 participants, including people from the Alan Turing Institute and The Turing Way and sessions at Open Data Day 2023 and Mozilla Festival 2023. All participants had positive impressions of "Am I in The Stack", but none of those interviewed said they wanted to opt out.[3] Developers who did opt out were asked why. Recurring answers included a preference for opt-in rather than opt-out consent, a view that using their code without compensation was unfair, worry about legal liability if generated code could be traced to their work, a belief that their code was too poor to train on, and PII in their code.[3] The governance card also acknowledges a structural limit of opt-out: permissively licensed code is often copied into other repositories, so an owner's opt-out does not remove every copy.[3]

### Personal information

To reduce the risk that a model would reproduce private data such as passwords or email addresses, BigCode redacted PII before training. The governance card says regular-expression tools and commercial PII software did not meet the project's requirements, so it built its own annotated dataset.[3] Through Toloka, 1,399 crowd workers in 35 countries annotated 12,000 code files containing 22,950 PII entities. Pay was set so that the hourly rate was above the US federal minimum wage and equivalent, in purchasing power, to the highest US state minimum wage.[3][9] The resulting detector, StarPII, is based on an encoder model the project trained on The Stack (StarEncoder); the governance card reports a 90% F1 score for it, higher than regex-based tools, especially for secret keys.[3][12] The annotated dataset itself is gated and shared with researchers case by case.[3]

### Attribution

With SantaCoder, and again with StarCoder, BigCode released search tools that let users check whether generated code appears in the pretraining data: a fast membership test and a full-text search that returns likely matches with their source.[3][12] The StarCoder paper describes a two-step design in its VSCode demo, a lightweight membership check followed by a search over a BM25 index.[9]

## Models

| Model | Released | Parameters | Training data | License |
|---|---|---|---|---|
| [SantaCoder](https://aiwiki.ai/wiki/santacoder) | December 22, 2022 | 1.1B | Python, Java and JavaScript from The Stack | CodeML OpenRAIL-M v0.1 |
| StarCoderBase | May 4, 2023 | 15.5B | 1 trillion tokens from 80+ languages, GitHub issues, Git commits and Jupyter notebooks | BigCode OpenRAIL-M v1 |
| [StarCoder](https://aiwiki.ai/wiki/starcoder) | May 4, 2023 | 15.5B | StarCoderBase further trained on 35B Python tokens | BigCode OpenRAIL-M v1 |
| StarCoder2-3B | February 28, 2024 | 3B | The Stack v2, 3+ trillion tokens | BigCode OpenRAIL-M v1 |
| StarCoder2-7B | February 28, 2024 | 7B | The Stack v2, 3.5+ trillion tokens | BigCode OpenRAIL-M v1 |
| StarCoder2-15B | February 28, 2024 | 15B | The Stack v2, 4+ trillion tokens | BigCode OpenRAIL-M v1 |

Sources: SantaCoder dates and license from the governance card;[3] StarCoder details from the StarCoder paper;[9] StarCoder2 details from the Hugging Face announcement.[15]

### SantaCoder

The first of two model development cycles ran in November and December 2022 and ended with SantaCoder, a 1.1-billion-parameter model trained on the Java, JavaScript and Python portions of The Stack.[3] Its technical report, "SantaCoder: don't reach for the stars!", described experiments to de-risk the model architecture and the data pipeline. It found that more aggressive near-deduplication helped and that keeping only files from repositories with five or more GitHub stars made results significantly worse. The best model outperformed the larger InCoder-6.7B and CodeGen-Multi-2.7B on the Java, JavaScript and Python parts of the MultiPL-E benchmark, both for left-to-right generation and for infilling.[8]

### StarCoder

The second cycle, from January to April 2023, scaled training up to 15.5-billion-parameter models trained on 1 trillion tokens.[3] StarCoder and StarCoderBase, announced on May 4, 2023,[11] have an 8,192-token context window, [fill-in-the-middle](https://aiwiki.ai/wiki/fill_in_the_middle) infilling, and [multi-query attention](https://aiwiki.ai/wiki/mqa) for fast large-batch inference.[9] The paper reported that StarCoderBase outperformed every open code LLM supporting multiple programming languages and matched or beat OpenAI's code-cushman-001, and that StarCoder could be prompted to reach 40% pass@1 on HumanEval.[9] The paper was published in Transactions on Machine Learning Research in December 2023.[9] Other models in the release included StarCoder+, StarCoderBase further trained on English web data, and the StarEncoder and StarPII models; tools included a VSCode extension, a playground, the StarChat chat demo and a "tech assistant" prompt.[12]

ServiceNow's 2023 announcement described StarCoder as "open-access, open-science, and open-governance" and quoted de Vries and von Werra as co-leads of BigCode.[10] According to the 2024 StarCoder2 announcement, ServiceNow built its text-to-code Now LLM on a specialized version of StarCoder, and Hugging Face used the model to build its StarChat assistant.[14]

### StarCoder2

StarCoder2, released on February 28, 2024, brought [NVIDIA](https://aiwiki.ai/wiki/nvidia) in as a third partner. ServiceNow trained the 3B model with its Fast LLM framework, Hugging Face trained the 7B model with its nanotron framework, and NVIDIA built the 15B model with [NVIDIA NeMo](https://aiwiki.ai/wiki/nvidia_nemo) and trained it on NVIDIA accelerated infrastructure.[14] The 3B and 7B models were trained on 17 programming languages and the 15B model on more than 600.[15] The paper reports that StarCoder2-3B outperformed the original StarCoderBase-15B and that StarCoder2-15B matched or outperformed CodeLlama-34B, a model more than twice its size, while DeepSeekCoder-33B remained stronger at code completion for high-resource languages.[13] At the launch, Software Heritage director Roberto Di Cosmo called StarCoder2 "the first code generation AI model developed using the Software Heritage source code archive".[14]

## Instruction tuning and alignment research

BigCode members have also published work on adapting base code models to follow instructions.

- **OctoPack** (August 2023) used Git commits, which pair code changes with human-written messages, as instruction data. It compiled CommitPack, 4 TB of commits across 350 programming languages, and instruction-tuned StarCoder into OctoCoder, which reached 46.2% pass@1 on HumanEval, the best result at the time among models not trained on OpenAI outputs. The paper also introduced HumanEvalPack, which extends HumanEval to code repair, code explanation and code synthesis in six languages.[16]
- **Astraios** (January 2024) compared seven tuning methods, including full fine-tuning and [parameter-efficient](https://aiwiki.ai/wiki/peft) methods, across 28 instruction-tuned OctoCoder models of four sizes up to 16 billion parameters. It found that full fine-tuning generally performed best and that [LoRA](https://aiwiki.ai/wiki/lora) usually offered the best trade-off between cost and performance.[17]
- **SelfCodeAlign** (October 2024, accepted to NeurIPS 2024) is a pipeline in which a model generates its own coding tasks, writes responses and tests, and keeps only responses that pass in a sandbox, without extensive human annotation or distillation from another model. The paper says the method led to StarCoder2-Instruct, which it describes as the first fully transparent, permissively licensed, self-aligned code LLM.[18]

## Evaluation

Evaluation was one of the project's founding priorities.[1] The **bigcode-evaluation-harness**, inspired by EleutherAI's [lm-evaluation-harness](https://aiwiki.ai/wiki/lm_evaluation_harness), is a framework for evaluating code generation models. Its tasks include HumanEval, HumanEval+, [MBPP](https://aiwiki.ai/wiki/mbpp), MBPP+, APPS and DS-1000, the HumanEvalPack benchmark from OctoPack, MultiPL-E's translation of HumanEval into 18 languages, program-aided math problems, and code-to-text tasks from CodeXGLUE.[21]

[BigCodeBench](https://aiwiki.ai/wiki/bigcodebench), introduced in June 2024, asks models to solve 1,140 practical Python tasks that require calling functions from 139 libraries across seven domains, with an average of 5.6 test cases per task. Its evaluation of 60 LLMs found scores of up to 60%, against human performance of 97%. The paper was accepted as an oral presentation at ICLR 2025.[19]

**BigCodeArena** (October 2025) is a human evaluation platform for code generation built on [Chatbot Arena](https://aiwiki.ai/wiki/lmsys_chatbot_arena) that runs the code models produce, so that voters can judge results as well as source text. The authors collected more than 14,000 conversation sessions across 10 LLMs, 10 languages and 8 execution environments, from which they identified more than 4,700 multi-turn samples with pairwise human preferences. From this data they built two further benchmarks: BigCodeReward, which tests how well [reward models](https://aiwiki.ai/wiki/reward_model) agree with human preferences, and AutoCodeArena, an automatic Elo-style benchmark. The authors report that proprietary models such as [GPT-5](https://aiwiki.ai/wiki/gpt-5), Claude Sonnet 4 and Claude Opus 4 still led in code generation among recently released models.[20]

## Licensing

BigCode's intellectual property policy sets a different license for each kind of output:[22]

| Output | License |
|---|---|
| Code (inbound contributions and released code) | Apache 2.0 |
| Documentation (specifications, guidelines, blog posts, papers) | Creative Commons Attribution 4.0 |
| Machine learning models and checkpoints | An Open & Responsible AI License ([OpenRAIL](https://aiwiki.ai/wiki/openrail)) |
| Datasets | Released when the project has the rights to do so, with data cards |

SantaCoder was released under the CodeML OpenRAIL-M v0.1 license, and StarCoder and StarCoder2 under the BigCode OpenRAIL-M v1 license.[3][15] The license allows royalty-free use, modification and redistribution, including for commercial purposes, but anyone who shares the model or a modified version must carry over its use restrictions (or a similar set) and keep or supply a model card.[3][23] BigCode's FAQ says the license evolved from the OpenRAIL licenses created for BigScience and added a new restriction against using the model to generate or spread malware. The same FAQ states that it "is not an open source license according to the Open Source Initiative definition", because it restricts how the model may be used.[23] The license text itself is available under CC BY 4.0 so that others can adopt or adapt it.[3]

## Timeline

| Date | Event |
|---|---|
| 2022-09-26 | BigCode project announced |
| 2022-10-27 | The Stack dataset and paper introduced |
| 2022-11-15 | "Am I in The Stack" tool and opt-out process launched |
| 2022-12-01 | The Stack v1.1 released with 358 programming languages |
| 2022-12-22 | SantaCoder released |
| 2023-03-20 | The Stack v1.2 announced with a simplified opt-out process |
| 2023-05-04 | StarCoder and StarCoderBase announced |
| 2023-08 | OctoPack and CommitPack published |
| 2023-12 | Governance card published; StarCoder paper published in TMLR |
| 2024-02-28 | StarCoder2 and The Stack v2 released |
| 2024-06 | BigCodeBench published |
| 2024-10 | SelfCodeAlign published |
| 2025-10 | BigCodeArena published |

Sources: project milestones page and governance card for 2022 to May 2023;[3][4] papers and announcements for later entries.[3][9][13][16][18][19][20]

## See also

- [Hugging Face](https://aiwiki.ai/wiki/hugging_face)
- [ServiceNow](https://aiwiki.ai/wiki/servicenow)
- [StarCoder](https://aiwiki.ai/wiki/starcoder)
- [The Stack](https://aiwiki.ai/wiki/the_stack)
- [OpenRAIL](https://aiwiki.ai/wiki/openrail)
- [AI code generation](https://aiwiki.ai/wiki/ai_code_generation)
- [Open-source AI](https://aiwiki.ai/wiki/open_source_ai)

## References

1. BigCode. "The Mission". bigcode-project.org, accessed November 2026. https://www.bigcode-project.org/docs/about/mission/
2. BigCode. "Organization". bigcode-project.org, accessed November 2026. https://www.bigcode-project.org/docs/about/organization/
3. Hughes, S., de Vries, H., Robinson, J., Muñoz Ferrandis, C., Ben Allal, L., von Werra, L., Ding, J., Paquet, S., Jernite, Y. "The BigCode Project Governance Card". arXiv:2312.03872, December 2023. https://arxiv.org/abs/2312.03872
4. BigCode. "Milestones". bigcode-project.org, accessed November 2026. https://www.bigcode-project.org/docs/about/timeline/
5. Wiggers, K. "Hugging Face and ServiceNow launch BigCode, a project to open source code-generating AI systems". TechCrunch, 2022-09-26. https://techcrunch.com/2022/09/26/hugging-face-and-servicenow-launch-bigcode-a-project-to-open-source-code-generating-ai-systems/
6. Kocetkov, D., Li, R., Ben Allal, L., et al. "The Stack: 3 TB of permissively licensed source code". arXiv:2211.15533, November 2022. https://arxiv.org/abs/2211.15533
7. BigCode. "Datasets". bigcode-project.org, accessed November 2026. https://www.bigcode-project.org/docs/about/the-stack/
8. Ben Allal, L., Li, R., Kocetkov, D., et al. "SantaCoder: don't reach for the stars!". arXiv:2301.03988, January 2023. https://arxiv.org/abs/2301.03988
9. Li, R., Ben Allal, L., Zi, Y., et al. "StarCoder: may the source be with you!". Transactions on Machine Learning Research, December 2023 (arXiv:2305.06161). https://arxiv.org/abs/2305.06161
10. ServiceNow. "ServiceNow and Hugging Face release StarCoder, an open-access large language model for code generation". ServiceNow news release, 2023-05-04. https://s205.q4cdn.com/537566246/files/doc_news/ServiceNow-and-Hugging-Face-release-StarCoder-an-open-access-large-language-model-for-code-generation-05-04-2023-traffic-2023.pdf
11. von Werra, L., Ben Allal, L. "StarCoder: A State-of-the-Art LLM for Code". Hugging Face blog, 2023-05-04. https://huggingface.co/blog/starcoder
12. BigCode. "Models". bigcode-project.org, accessed November 2026. https://www.bigcode-project.org/docs/about/models/
13. Lozhkov, A., Li, R., Ben Allal, L., et al. "StarCoder 2 and The Stack v2: The Next Generation". arXiv:2402.19173, February 2024. https://arxiv.org/abs/2402.19173
14. ServiceNow, Hugging Face and NVIDIA. "ServiceNow, Hugging Face, and NVIDIA launch New Open-Access LLMs Help Developers Tap Generative AI to Build Enterprise Applications". ServiceNow news release, 2024-02-28. https://s205.q4cdn.com/537566246/files/doc_news/ServiceNow-Hugging-Face-and-NVIDIA-launch-New-Open-Access-LLMs-Help-Developers-Tap-Generative-AI-to-Build-Enterprise-Applications-02--XYIO6.pdf
15. von Werra, L., Ben Allal, L., Lozhkov, A., Tazi, N. "StarCoder2 and The Stack v2". Hugging Face blog, 2024-02-28. https://huggingface.co/blog/starcoder2
16. Muennighoff, N., Liu, Q., Zebaze, A., et al. "OctoPack: Instruction Tuning Code Large Language Models". arXiv:2308.07124, August 2023. https://arxiv.org/abs/2308.07124
17. Zhuo, T. Y., Zebaze, A., Suppattarachai, N., et al. "Astraios: Parameter-Efficient Instruction Tuning Code Large Language Models". arXiv:2401.00788, January 2024. https://arxiv.org/abs/2401.00788
18. Wei, Y., Cassano, F., Liu, J., et al. "SelfCodeAlign: Self-Alignment for Code Generation". NeurIPS 2024 (arXiv:2410.24198), October 2024. https://arxiv.org/abs/2410.24198
19. Zhuo, T. Y., Vu, M. C., Chim, J., et al. "BigCodeBench: Benchmarking Code Generation with Diverse Function Calls and Complex Instructions". ICLR 2025 (arXiv:2406.15877), June 2024. https://arxiv.org/abs/2406.15877
20. Zhuo, T. Y., Jin, X., Liu, H., et al. "BigCodeArena: Unveiling More Reliable Human Preferences in Code Generation via Execution". arXiv:2510.08697, October 2025. https://arxiv.org/abs/2510.08697
21. BigCode. "bigcode-evaluation-harness" (README). GitHub, accessed November 2026. https://github.com/bigcode-project/bigcode-evaluation-harness
22. BigCode. "How we manage IP". bigcode-project.org, accessed November 2026. https://www.bigcode-project.org/docs/about/ip/
23. BigCode. "BigCode OpenRAIL-M" (license FAQ). bigcode-project.org, accessed November 2026. https://www.bigcode-project.org/docs/pages/bigcode-openrail/
24. Lozhkov, A., Larcher, H., Morlon, M., Ben Allal, L., von Werra, L. "The Stack v3: The Largest Open Code Dataset". Hugging Face, 2026. https://huggingface.co/datasets/HuggingFaceCode/stack-v3-train
25. BigCode. "opt-out-v2: Repository for opt-out requests". GitHub, accessed November 2026. https://github.com/bigcode-project/opt-out-v2

