Fermatix AI — End-to-end data solutions for code LLMs and AI agents

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. Every handoff is reproducible.

Data sourcing

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

Generation & SFT

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

Annotation & review

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

Benchmarks & harnesses

SWE-Bench, Multi-SWE, Terminal-Bench, RAG eval, Dockerized reproducible environments.

Evaluation & red team

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

Datasets

Ready-to-license datasets

Non-public code repositories, SWE-Bench benchmarks, alignment sets, and enterprise data.

01 Non-Public Code Repositories for LLM Pre-Training

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

Around 3,000 proprietary codebases that have never 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.

Snapshot

~3,000

Repositories

300M+

Lines of code

100K+

Commits

500K+

Files

~25

Avg contributors

2018–25

Creation period

Distribution: JS/TS 35%, PHP 30%, Obj-C/Swift 12%, Java 8%, Python 4%, Other 11%.

Composition: 54% discontinued / 46% active or maintained. Full legal rights to license every repository.

Why this matters

Zero contamination risk. None of this code exists in public datasets — cleaner pre-training signal, more reliable benchmark evaluation.
Human-authored guarantee. All code reviewed by the engineering team. For post-2024 code, any LLM-assisted portions are redacted and verified as production-grade, human-reviewed output.

Start with a curated pilot subset to validate quality and fit before scaling.

02 SWE-Bench-Style Benchmarks

Fully compatible with the Multi-SWE-Bench framework.

Open-Source SWE-Bench

~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

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.

A ready-made non-public Python repository is available as a pilot — fully annotated and bench-ready.

Coverage by language

Kotlin327

Rust250

Go207

PHP166

C++122

Java110

C109

C#102

Ruby80

TS47

Scala39

JS18

Total1,577

→ View sample (Hugging Face)

03 Post-training data

Fine-tuning, alignment, RAG and safety data for code models.

SFT dialogues — instruction-tuning data authored by practicing engineers, not crowd workers.
RLHF & DPO pairs — pairwise preferences, rejection sampling across custom rubrics.
Prompt-to-Code datasets — generation-focused training data.
RAG evaluation — retrieval-augmented generation evaluation datasets.
Red teaming — adversarial multi-step scenario datasets.
Open-source dataset expansion — popular open datasets extended with new sources, domains, and refined taxonomy.

04 Agent data

Trajectories, architecture eval, dialog scoring and safety probes for code agents.

Agent behavior trajectories — datasets of agent actions and reasoning for accuracy and performance improvements.
Agent architecture analysis — evaluation datasets for architectures: search, integration, tools, and external systems.
Dialog evaluation — datasets scoring dialog history across multiple criteria.
Dialog safety — evaluation suites for politeness, honesty, and integrity.

05 Enterprise Data

Internal enterprise data sourced from real companies — active, acquired, or wound down — each with certified consent to license.

Data types

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

Coverage & compliance

~20

Companies now

100+

Scalable to

Certified consent from each source company.
Full PII redaction.

Start with a 10-company pilot at a bulk rate.

→ View metadata (Google Sheets)

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, no glue code on your side.

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.

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

Track record

Engineering team & domain experts

Multi-level QC

GDPR · ISO