github.com/brocode
Every commit linked. Every dataset cited. Every benchmark reproducible.
The Open-Source Hour is paid time for every Brocode engineer to contribute upstream — and the contribution graph below is the result.
brocode org — last 12 months
312
merged PRs
14
upstream projects
7
public datasets
1
Khaleeji Benchmark
Open-Source Hour: every engineer, every Friday afternoon, since 2023.
Why we publish
Signed by the CTO and the head of Arabic NLP.
Open source is the only signal a buyer or a researcher cannot fake. Code that runs in public gets reviewed by people without commercial interest in agreeing with you. That is the honest test of an engineering practice, and it is the test we built the Open-Source Hour around — paid time, every Friday afternoon, for every engineer in the company, audited quarterly by the CTO.
For buyers doing tie-breaker due diligence between two shortlisted vendors, the contribution graph and the merged-PR ledger answer the engineering-culture question without a single marketing slide. For researchers evaluating Brocode as a future employer or research collaborator, the same artefacts answer the career-relevance question.
Repositories and contributions
The named projects, the upstream PRs, the dataset cards.
Benchmark
Pythonbrocode/khaleeji-benchmark
Open evaluation suite for Arabic LLMs covering UAE / KSA / Qatar / Kuwait / Bahrain dialect comprehension, MSA reasoning, and Arabic-English code-switching. Public leaderboard with Falcon, Jais, Claude, GPT-4 and Brocode fine-tunes.
★ 1.8k
Dataset
Datasetbrocode/khaleeji-dialect-corpus-v2
1.4M utterances across UAE, KSA, Qatar, Kuwait and Bahrain dialect markers. Apache-2.0 licensed, dataset card with provenance, statistics and ethics review.
★ 3.2k
Dataset
Datasetbrocode/uae-government-correspondence-ner
Anonymised NER tags on Arabic correspondence templates, designed for the long-form formal Arabic that dominates UAE government workflows.
★ 0.9k
Dataset
Datasetbrocode/arabic-financial-extraction-eval
1,800 labelled Arabic invoices and KYC documents for extraction benchmarking. Released with category breakdowns and a reproducer notebook.
★ 0.7k
Upstream PR
Rusthuggingface/tokenizers — Arabic normalisation
Upstream PR fixing ZWNJ, tatweel, and alef-variant handling in the Hugging Face tokenizers package. Merged into main; ships from 0.20.0.
★ merged
Upstream PR
Pythonvllm-project/vllm — RTL paged-attention fixes
Paged-attention edge-case fixes for right-to-left scripts, Arabic-locale tokenisation pathway, and sample serving configs for Falcon-family and Jais models.
★ merged
Upstream PR
PythonEleutherAI/lm-evaluation-harness — Arabic task pack
Public PR adding seven Arabic MCQA and generation tasks aligned to the Khaleeji Benchmark, with reproducer scripts and dataset cards.
★ merged
Upstream PR
Cpgvector/pgvector — high-dim build perf
Index-build performance fix on high-dimension Arabic-embedding workloads. Ships from 0.7.4 onwards.
★ merged
312
Merged PRs across 14 upstream projects
7
Public Arabic datasets (4.6M records)
1
Open Arabic-LLM benchmark with leaderboard
Fri
Open-Source Hour — every engineer, every week
The Khaleeji Benchmark
An open Arabic LLM benchmark with a public leaderboard.
UAE / KSA / Qatar / Kuwait / Bahrain dialect comprehension, MSA reasoning, Arabic-English code-switching. Refreshed within 14 days of any major frontier-model release.
| Model | Dialect | MSA reasoning | Code-switch |
|---|---|---|---|
| Claude 3.5 Sonnet | 74.1 | 82.4 | 79.6 |
| GPT-4o | 71.3 | 83.9 | 78.2 |
| Brocode-Jais-FT-v3 | 78.6 | 76.1 | 74.4 |
| Falcon-3-7B-instruct | 64.5 | 69.7 | 66.0 |
| Jais-30b-chat-v3 | 67.2 | 71.8 | 68.9 |
Illustrative snapshot from the public leaderboard; live numbers update on each release.
Versus the alternatives
What an engineering-culture due-diligence comparison actually looks like.
| Capability | Brocode | Offshore consultancies | Big-4 AI practice | Sovereign-only integrator |
|---|---|---|---|---|
| Public commit graph for the consultancy org | Yes — Brocode org + named maintainer handles | No | No | Limited internal repos |
| Arabic NLP datasets on Hugging Face | 7 datasets, 4.6M labelled records | 0 | 0 | 1–2, internal-only |
| Merged PRs to vLLM / Transformers / tokenizers / pgvector | 312 merged across 14 projects | 0–5 | 0 | <20 |
| Open Arabic LLM benchmark with leaderboard | Khaleeji Benchmark — public leaderboard | No | No | Internal moat |
| Paid Open-Source Hour for every engineer | Yes — every Friday afternoon since 2023 | No | No | No |
| Named maintainer attribution | Yes — engineers commit under their own GitHub handles | Anonymised | Anonymised | Anonymised |
The Open-Source Hour policy
Paid time, every Friday afternoon, audited by the CTO.
It is a clause in our employment contract. It is a line in our delivery rate card. It is the reason the contribution graph above is real.
policy.1
Every engineer, every Friday afternoon
No exceptions for delivery pressure. Four hours of paid time on Friday afternoon for upstream work; if a sprint demands the time, the sprint plan is wrong.
policy.2
Audited by the CTO every quarter
Per-engineer contribution counts are reviewed by the CTO; zero-contribution quarters trigger a conversation, not a sanction.
policy.3
Public attribution by consent
Engineers commit under their own GitHub handles with their consent — never aggregated into a single corporate handle that obscures who did the work.
Free download
Brocode AI Open-Source Report 2026
The 36-page PDF with the full commit ledger by repository, every dataset card with provenance and statistics, the Khaleeji Benchmark methodology, and a public bibliography.
- Khaleeji Benchmark (open dataset + leaderboard)
- Arabic NLP datasets — 7 datasets, 4.6M labelled records
- vLLM contributions — RTL paged-attention, Arabic-locale path
- tokenizers contributions — ZWNJ, tatweel, alef-variant fixes
- pgvector contributions — high-dim index-build perf fix
- Internal tools we open-sourced
- Open-Source Hour audit ledger
Engineering-culture questions
What buyers and researchers ask before a maintainer call.
The contribution graph spans 18 months of activity with dense weeks across every quarter, driven by the Open-Source Hour policy that pays every engineer for one Friday afternoon a week to contribute upstream. The report contains the per-quarter contributor count, PR-merge cadence and dataset-release schedule — none of it is a marketing burst.
Talk to a maintainer
Arabic NLP, serving infra, data engineering, or MLOps.
The form routes directly to the maintainer specialism you choose. One business day to first response.
Researcher or engineer interested in joining? See Brocode engineering careers.
Prefer chat? Message us on WhatsApp.
Continue exploring
Related capabilities and stories
Technology Stack
Every contribution maps back to a component in the stack.
Read moreNatural Language Processing
The commercial reflection of the Arabic NLP work.
Read moreGenerative AI Development
Where the vLLM and serving contributions show up.
Read moreQuality Assurance for AI
The evaluation harness behind the Khaleeji Benchmark.
Read moreCareers — engineering
Where the Open-Source Hour policy lives in the contract.
Read more