End‑to‑end data solutions for code LLMs and AI agents

Fast, no‑friction execution across every stage of code‑data collection and annotation — from pre‑training to post‑training.

4,200+
Licensed repositories
322K
Datapoints delivered
21
Programming languages
30
Engineers & experts
Team

Engineering team & domain experts

30 engineers, annotators, and domain experts. Rigorous hiring pipeline — up to 100 structured interviews per week.

Quality

Multi‑level QC

Cross‑validation by practicing industry experts. Multi‑stage QC gates and per‑rater calibration.

Compliance

GDPR · ISO

Compliance with data security and confidentiality standards. Full PII redaction across enterprise data.

How we work

The full lifecycle, end to end

Not just a data supplier — we own every stage from sourcing to evaluation. Plug us in at any step, or hand off the whole chain.

01

Data sourcing

Non‑public repos, real enterprise content, licensed archives — never seen in public sets.

02

Generation & SFT

Prompt → response pairs, multi‑turn dialogues, code‑task authoring at scale.

03

Annotation & review

Human‑in‑the‑loop labeling, rationales, pairwise ranking across 40 criteria.

04

Benchmarks & harnesses

SWE‑Bench, Multi‑SWE, Harbor, RAG eval, Dockerized reproducible environments.

05

Evaluation & red team

Agent‑trajectory scoring, plan/thought eval, safety probes, regression testing.

Datasets

Ready‑to‑license datasets

What we have on the shelf, and what we build to order. Numbers 01 to 04 can be licensed today; 05 and 06 are produced against your brief.

01 Licensed Private Repositories

Production code from real companies — never indexed, never crawled.

4,200+ proprietary repositories from 1,900+ companies, 800M lines of code — none of it ever appeared in public training sets (GitHub, GitLab, HuggingFace). Production‑grade repositories from real companies, primarily sourced from a network of outsourcing agencies and startups whose products were discontinued or acquired. Projects written mostly by LLMs, very small repositories and forks are filtered out.

Snapshot

4,200+
Repositories
1,900+
Companies of origin
800M
Lines of code
38B
Characters
4.5M
Source files
1.6M
Commits
300K+
Pull requests
2010–26
Creation period

Language mix by repository

PHP 19% TypeScript 16% JavaScript 15% Python 11% Swift 6% Java 5% Kotlin 5% Other 23%

44 primary languages in all, with Dart, C#, Go, Rust and C/C++ making up much of the remainder. Each repository arrives with its language stack, SPDX‑compliant license type, creation date, project category, project status and line count, and with its commit history.

Composition: 67% discontinued / 33% active or maintained. Full legal rights to license every repository.

Get catalog & samples → Collection methodology — arXiv:2605.12153 →

02 SWE‑Bench‑Style Benchmarks

1,577 public tasks across 12 languages, plus a 50+ task set built from the private repositories in 01.

Both categories are fully compatible with the Multi‑SWE‑Bench framework, and every task is reviewed by a developer.

Open‑Source SWE‑Bench

  • 1,577 tasks across 12 languages (~8,712 files, ~8.85 GB)
  • Task types: bug fixing, code completion, PR generation, automated code review, regression validation, environment‑setup verification
  • Artifacts: issue descriptions, PRs, commit messages, golden/test patches, install scripts, Dockerized environments, Parquet metadata
  • Verified: every task tested by real developers

Non‑Public‑Repository SWE‑Bench

  • 50+ tasks built from the licensed private repositories in 01.
  • True out‑of‑distribution evaluation. Models can't memorize these tasks.
  • Multi‑language support (not Python‑only like public SWE‑Bench Verified).
  • Golden/test patches, Dockerized environments, Parquet metadata for reproducible patch‑apply, build, and test workflows.

Coverage by language

Kotlin327
Rust250
Go207
PHP166
C++122
Java110
C109
C#102
Ruby80
TS47
Scala39
JS18
Total1,577

Get metadata & samples → Public sample on Hugging Face →

03 Terminal‑Bench‑Style Tasks

100 self‑contained terminal and development‑environment tasks, calibrated so the latest Opus and GPT pass under 50%.

100
Tasks
<50%
Pass rate, latest Opus and GPT

What each task contains

  • An instruction, written the way a developer would receive it
  • A Docker environment the agent works inside
  • A verifier that decides pass or fail
  • A gold solution to score against

Why a run is comparable

Delivered in Harbor format. Public base images are pinned by digest and nothing reaches the network while a task runs, so the same task scores the same way on our machines and on yours. Each trial is recorded with the model, the harness version and the reward it earned.

Get samples →

04 Enterprise Workflow Dataset

Documents, tracker tasks, messages, meetings and call transcripts, with the links between them intact.

All of it comes out of company‑internal systems whose owners gave certified consent. You can follow one piece of work across every system it touched, which is what an enterprise agent has to do when the work does not fit into a single exchange.

Data types

  • Team communication platforms (Slack, Discord, CRM systems)
  • Task trackers (Jira, YouTrack)
  • Meeting recordings with transcripts
  • Knowledge bases and internal documentation

What it captures is the ordinary texture of a company at work: how tasks get planned, how technical arguments play out, how decisions land, and how teams talk to each other across functions.

One company, built out end to end

24,400
Files
96M
Words (~325M tokens)
1,300
Hours of calls

Coverage & compliance

  • The same collection runs across several dozen companies.
  • Certified consent from each source company.
  • Full PII redaction and GDPR‑aligned handling.

Get metadata & samples →

05 Custom Datasets

Built to order from your repositories, trackers and internal workflows.

Experts turn each task into a clean starting state, an expected outcome and a validation procedure, so the result works for training or for evaluation. Every item is reviewed independently, and where the task allows it we check the result against tests, golden patches or scoring rules agreed with you up front.

322K
Datapoints delivered
50+
Custom datasets
21
Programming languages
6+
Clients

Coding

  • Agent trajectories — read, edit, run tests, debug, fix, over real repositories with verifiable outcomes
  • SFT and RL tasks with checked outputs, plus issue and merge request pairs
  • Issue resolution from held‑out repositories, so the set resists contamination
  • Debugging and test repair — failing test, root cause, fix
  • Code review and pairwise preference for RLHF and reward modelling

General

  • Domain‑expert data — Q&A and reasoning traces from credentialed experts in STEM, medical, legal and finance
  • Evaluation — rubric scoring, LLM‑judge pipelines, private benchmarks with executable checks
  • Safety and red teaming — adversarial prompts, refusals, jailbreak resistance

Request task catalog & examples →

06 Search‑Grounded Synthetic Data

Everything above is real data. This is the one entry that is generated.

We can help on the synthetic side too, and the way we do it came out of an experiment. We took the BeyondWeb idea and put a search engine underneath it: rather than paraphrasing a source document in isolation, an agent decides which claims matter and what context is missing, retrieves independent evidence for them, then writes a synthesis across those documents. Every example keeps its source trail, so you can see where a claim came from.

58:0
Wins against Nemotron‑Synth
18:0
Search vs no search
8.59
Mean judge score, vs 7.37 and 5.55

We can point the same pipeline at a model architecture, an evaluation rubric, or a slice your existing dataset is missing.

Request experiment report →

Integration

Data delivered in the shape your pipeline expects

Not just data delivery — our engineers integrate at every stage of your pipeline. Standard formats, your evaluation frameworks, schema design, ingestion adapters, continuous delivery on your cadence — data arrives ready to train or benchmark.

Formats

Drop‑in formats and schemas

JSONL for SFT and DPO, Parquet for pre‑training corpora, HuggingFace Datasets for publishing. Conversation data in ShareGPT, Alpaca, or OpenAI chat schemas — drop‑in for your training loop, no conversion on your side.

Reproducibility

Reproducible eval environments

Dockerized SWE‑Bench and Multi‑SWE‑Bench harnesses (Harbor‑compatible), RAG eval — patch‑apply, build, and test runs work identically on our machines and yours.

Embedded experts

Engineers inside your roadmap

Practicing developers, ML engineers, and domain specialists as an extension of your team — scoping taxonomies, defining quality criteria, resolving edge cases as they surface.

Blog

Research & engineering notes

FAQ

Questions we get before the first pilot

Who builds the data, how a first project runs, and the two things every team asks about the non‑public repository collection.

Why should we choose Fermatix?

We came to AI data from software development, not the other way round. The company wrote production code first and started its ML and AI work in 2017. Today we work with frontier labs on coding data and model evaluation.

Who builds the datasets, and how we scale them →

Do you use crowdsourcing?

No. Every coding task is written and reviewed by our own middle and senior engineers. We match the work to background, so an ML task goes to an ML engineer and an Android task to someone who ships Android.

Do you offer ready‑made data or custom projects?

Both. You can license data and benchmark tasks we already have, or commission a set built around a specific language, skill, format or difficulty level. The cheaper order is usually to test an existing sample first: once we can see where your model actually struggles, the custom work goes where it pays off.

How do you evaluate quality?

The rubric and the baseline are agreed before production starts, never after. Depending on the task, checks can include automated tests, model-based review, blind comparison, review by a developer, and metrics from the model downstream. We normally start with a POC, so you can weigh quality, cost and the effect on the model before committing to a bigger run.

What is included in your private repository dataset?

Whole codebases from real companies, with the project structure and commit history included. None of it was ever published as open source. The languages run to JavaScript, TypeScript, PHP, Python, Objective‑C and Java among others, and the products behind them come from banking, retail, healthcare, logistics, education and IT services.

Do you have the right to license the repositories?

Yes. Every repository comes to us under an agreement that lets us license the data, and the rights holders are paid royalties on it. We can show the supporting paperwork under NDA. Ownership of the original code, our right to license it, and any exclusivity you want are three separate things, and the contract spells out each one.

All questions →

Contact

Start a pilot

Curated subset of non‑public repositories and benchmark tasks for hands‑on quality validation. Schedule a technical deep dive with our engineering team.

Email: hi@fermatix.ai


AVENIDAS INTELIGENTES, LDA

Lg Alberto Sampaio, 3 A, Sala 10

Linda a Velha, 2795‑007

Portugal