Skip to content
trainjustby appx
Request an audit
Backed byY Combinator

A model is only as good as what it was trained on.

Trainjust builds expert-verified training data, annotation pipelines, and evaluation suites for production AI — anchored in 3.6M+ hours of multilingual audio, 2.5M+ patient records, textbooks in 15 languages, and 12K+ codebases.

Tell us about your data needsExplore the corpus
post-training stackmeasuredAudio3.6M+ hrsTextbooks50K+ booksClinical2.5M+ recordsCode12K+ reposCurationdedupe · factuality score · PII redactionISBN · language · vertical · channelPre-training15 langs · scoredSFTreasoning tracesRLHFexpert preferenceEvaluationbenchmark delta · factuality % · eval pass rategold-set validators · three-tier QA · IAA trackedProduction model01 / SOURCE02 / CURATE03 / TRAIN04 / EVALUATE
  • Arabic
  • Bahasa
  • Swahili
  • French
  • Nepali
  • Urdu
  • Bengali
  • Portuguese
  • Hindi
  • Kinyarwanda
  • Luganda
  • Java
  • Python
  • C++
  • JavaScript
  • C#
3.6M+
Total audio hours
Call center and podcast coverage
50K+
Total books
3B+ words across 15 languages
9.2M+
Total Q&A
1.8B+ tokens with explanations
4.4M+
Total artifacts
CAD · SVG · PSD · Infographics
4.7B+
total code tokens
Includes 25M+ curated DSA tokens.

Client portfolio

We've worked with some of the world's leading companies

  • Google
  • Amazon
  • AppX
  • Tencent
  • Meta
  • Webinar.gg

01The problem

Models can talk. That is not the same as being reliable.

Most production failures — hallucinations, shallow reasoning, misaligned rewards — start in the training data, not the architecture. The fix is earlier in the stack than most teams look.

01Hallucination at scale
Confident, fluent, wrong. Without factuality controls in the training mix, failures stay cheap to miss and expensive once they hit a regulated workflow.
02Brittle reasoning
SFT without explicit reasoning traces teaches pattern-matched answers. The model sounds right without doing the work that would make it right.
03Misaligned reward signals
RLHF from commodity crowdwork rewards sounding preferred, not being correct. The model learns polish over truth.

02The solution

Build the training signal before you ship the model.

Trainjust combines a proprietary multilingual corpus, domain experts, structured annotation, and evaluation design so AI teams can control what models know, how they reason, and where they break.

Each engagement starts with a quality baseline. Deliverables are scored against it — benchmark deltas, factuality lifts, and eval pass rates.

CorpusPre-training data curation
ISBN-attributed textbooks · 15 languages · factuality-scored · not web-scraped.
Expert-createdSFT dataset design
Instruction datasets with explicit reasoning traces, written by domain subject-matter experts.
Human-verifiedRLHF & reward modeling
2.1M+ hours of real industry audio · dual-channel · industry-tagged · WER-tracked.
HealthcareMedical AI training data
68.5M+ DICOM images · 1.6M+ patient records · 20 specialties · age-stratified.
CodingCode & reasoning datasets
4.7B+ total code tokens · 25M+ curated DSA tokens · 7 languages · curated, not scraped.
MeasurableFactuality auditing
Systematic identification of hallucination-prone examples, with before-and-after documentation.

03Corpus

See the training data before it enters the model.

Fourteen collections — audio, video, healthcare, textbooks, Q&A, coding, visual rendering, egocentric video, theses, legal contracts, longitudinal time series, operations, AI trust and safety, and meeting audio. Expert-curated, modality-tagged, licensed for enterprise training.

Index of collections

Audio corpus

3.6M+ hours of multilingual audio

Call center, podcast, and speech data with metadata for gender, age, industry, channel, dialect, and language.

Request an audio sample
3.6M+
call center hours
57.5K+
podcast hours
10+
podcast languages
50+
call center languages
4
audio refining steps
Dual
channel support

Call Center Audio - 50+ Languages

Single + Dual Channel

English · Arabic · French · Swahili · Bahasa · Urdu · Portuguese · Kinyarwanda · Luganda · Nepali · Bengali · and 35+ more

Call center audio hours by language
LanguageCodeCoverageHoursChannel
English (US)en_US52%144.2K hrsDual
English (UK)en_GB45%101.4K hrsDual
English (India)en_IN59%200K hrsDual
Arabicar_XA34%80K hrsDual
Frenchfr_FR35%41.6K hrsDual
Swahilisw_KE48%201.6K hrsDual
Bahasa Indonesiaid_ID68%315K hrsDual
Punjabipa_IN35%378K hrsDual
Marathimr_IN37%78K hrsDual
Lugandalg_UG33%92K hrsDual
Portuguesept_BR41%78.2K hrsDual
Nepaline_NP72%297K hrsDual
Bengalibn_BD89%385K hrsDual
+ 28 more languages—100%Hindi (1.6M+), Marathi, Assamese, Odia, Mizo, Chichewa, Ganda, Tamil, Gujarati...—

Industry verticals

  • Banking
  • Healthcare
  • Insurance
  • Agriculture
  • Telecom
  • QSR
  • Logistics
  • Humanitarian Aid
  • SaaS
  • Real Estate

Key USP - Dual Channel

Agent and Customer on separate tracks. Gold standard for speaker diarization training. No public dataset matches this at scale.

Podcast Audio - 10+ Languages

57.5K+ Total Hours

Natural long-form conversational speech across global and regional languages. Covers Arabic, South Asian, and Southeast Asian languages. Ideal for audio-LLM cross-modal alignment where organic multilingual discourse matters.

[ Podcast hours ]
  1. Arabic6K hrs
  2. Urdu2.3K hrs
  3. Bengali7.8K hrs
  4. Telugu6K hrs
  5. Punjabi4.8K hrs
  6. Marathi4.2K hrs
  7. Kannada4K hrs
  8. Malayalam4K hrs
  9. Tamil3.3K hrs
  10. Gujarati2.5K hrs

04Dataset USPs

What makes these datasets structurally different.

Benchmarked against major providers. Each claim ties back to files in the delivered corpus.

USP 01

Real Spontaneous Speech - Not Scripted

Live call center conversations capture disfluencies, interruptions, code-switching, and domain vocabulary used organically. Models trained on real conversations generalize far better to production than prompted audio.

vs. acted / prompted: more realistic acoustic and lexical variation

USP 02

Dual Channel - Agent & Customer Separate Tracks

Perfect diarization ground truth without post-hoc separation. Unlocks contact center AI, agent assist, compliance monitoring. Available for English (US/UK/IN), French, Arabic, Nepali, Bengali, Swahili, and more.

Gold standard for speaker diarization - no public dataset matches this at scale

USP 03

15+ Industry Verticals - Native Domain Vocabulary

Banking, healthcare, insurance, agriculture, telecom, QSR, humanitarian aid, SaaS, real estate, solar energy, travel. Industry-specific vocabulary inherent in the data - not annotated after the fact.

Industries: Banking · Healthcare · Insurance · Agriculture · Telecom · QSR · SaaS · Logistics · Humanitarian Aid · Real Estate

USP 04

Low-Resource Global Languages Unavailable Elsewhere

African languages: Swahili (146.3K hrs), Kinyarwanda (3.3K hrs), Luganda (21.6K hrs), Ganda, Chichewa. South Asian: Nepali (235K hrs), Bengali (377K hrs), Odia (12.7K hrs). Southeast Asian: Bahasa Indonesia, Javanese. All with enterprise-grade quality annotation and dual-channel where available.

Swahili 146.3K hrs · Nepali 235K hrs · Kinyarwanda 3.3K hrs · no commercial alternative at this depth

USP 05

Multilingual Code-Switching - Natural, Not Scripted

Real-world multilingual environments naturally blend languages within conversations - Arabic-French in North Africa, Bahasa-English across Southeast Asia, English-local language in enterprise call centers globally. Trainjust's corpus captures this cross-lingual mixing organically at scale, enabling voice AI models that handle real-world multilingual conversation without brittle hard language boundaries.

No synthetic or scripted dataset replicates the natural code-switching patterns found in real enterprise call centers

USP 06

Enterprise Data Agreements - Consent-Chain Documented

Collected via formal agreements with enterprise operators. EU AI Act Article 10, India DPDP Act, and CCPA require documented provenance. Trainjust's collection architecture is built for this from the ground up.

Not web-scraped · GDPR-eligible lineage · enterprise agreement structure

05Capabilities

One vendor from pre-training through evaluation.

Pre-training, SFT, RLHF, domain corpora, and eval — one quality bar, no handoffs between shops.

  1. 01Pre-training curation

    ISBN-attributed textbooks · 15 languages · deduplicated · not web-scraped. FineWeb-Edu achieved +24% ARC from educational text; ours comes from published books.

  2. 02SFT dataset design

    Instruction datasets with explicit reasoning traces. Built by domain SMEs (attorneys, physicians, researchers, engineers), not crowdworkers. Chain-of-thought citations included.

  3. 03RLHF & reward modeling

    2.1M+ hours of real industry audio · dual-channel · 35+ languages · WER-tracked. Speaker-disjoint splits — the preference signal reward models need to calibrate.

  4. 04Medical AI training data

    68.5M+ DICOM images paired with radiology reports · 1.6M+ patient records · age-stratified disease taxonomy · 20 specialties · real Indian hospital data.

  5. 05Code & reasoning datasets

    4.7B+ total code tokens · 25M+ curated DSA tokens · 7 languages · cross-language parallel structure. Same problem in Java, Python, C++, JavaScript, and C# — curated, not scraped.

  6. 06The Trainjust corpus

    Audio · Textbooks · Healthcare · Coding. Four flagship collections across 50+ call center languages, 15 textbook languages, and 7 programming languages. Expert-curated, enterprise-licensed, ready to train.

06Why Trainjust

What most data vendors leave out of the stack.

  1. 01

    A corpus you reuse, not a one-off delivery.

    2.1M+ audio hours, 1.6M+ patient records, a 15-language textbook library, and 64K+ coding solutions grew across years of production work. Each engagement adds to a retained, quality-scored asset. You start from verified examples in your domain instead of an empty bucket.

    Coverage: audio, healthcare, textbooks, coding, Q&A, image, video, and egocentric data with reviewable metadata.

  2. 02

    Every example points back to its source.

    Textbooks carry ISBNs. Audio hours carry language code, industry vertical, and channel type. Medical records carry modality, specialty, and clinical fields. Code problems carry language and algorithm category. Traceability is how training data stays trustworthy — compliance paperwork is the side effect.

    EU AI Act Article 10 requires documented training data sources for high-risk AI. This corpus is structured for that.

  3. 03

    Score the model, not the file count.

    Success is benchmark deltas, factuality lifts, and eval pass rates. FineWeb-Edu showed educational text beats web scrapes by +24% ARC and +12% MMLU. Our textbooks arrive educational by source. Our audio grounds speech models in ways open sets don't. Our medical set is the only Indian hospital corpus at this volume with paired DICOM and PDF.

    FineWeb-Edu, NeurIPS 2024: educational text → +24% ARC, +12% MMLU. Trainjust textbooks: educational by source.

07Use cases

For deployments where a wrong answer has a cost.

08Global presence

Annotation ops across two continents.

Offices in San Francisco and Noida, with language coverage across major commercial AI markets. That footprint supports enterprise SLAs, regional residency rules, and 24-hour delivery windows.

Office locations
CityCountryRoleCoordinates
San FranciscoUSAHeadquarters37.77° N, 122.42° W
NoidaIndiaCapability center28.58° N, 77.33° E

Headquarters

Capability center

09Infrastructure & R&D

Annotation is the interface. Infrastructure is the product.

Purpose-built annotation platforms, GPU processing pipelines, and quality tooling. The team includes NLP researchers, computational linguists, ML engineers, and domain experts — capacity a crowdsourcing marketplace does not ship.

01Compute

GPU-accelerated processing
Dedicated GPU infrastructure for ASR model runs, OCR pipelines, speaker diarization, MinHash deduplication at scale, and language classification across 176+ languages. Processing that would take weeks on CPU runs in hours.

02Tooling

Proprietary annotation platform
Purpose-built annotation interfaces for each data type: audio waveform editors with word-level alignment review, DICOM annotation tools with radiology report linking, and structured Q&A curation workflows with expert verification steps.

03Research

Corpus research & methodology
A dedicated R&D function benchmarking against FineWeb-Edu, Dolma, Nemotron-CC, and production ASR baselines. Methodology papers, quality evaluation frameworks, and benchmark datasets published openly.

04QA systems

Multi-tier quality evaluation
Automated quality gates (SNR, WER, perplexity, duplicate detection) run before any human review. Human QA operates in three tiers: general annotators, senior domain reviewers, and expert gold-set validators. IAA tracked per language, per dataset.

05Compliance

AI Act & data governance
EU AI Act Article 10 documentation architecture built in. ISO-aligned data management processes. PII detection and redaction pipelines. Data lineage tracking from raw source to delivered training file, with an audit trail available for regulated clients.

06Expertise

Domain expert network
12,000+ verified domain experts across legal, medical, financial, and engineering verticals. Computational linguists for 40+ languages. Radiologists, pathologists, and clinicians for healthcare annotation. Senior software engineers for code curation and review.

10Frequently asked

What buyers ask before they sign.

01Can we license specific language subsets or domain slices?

Yes. The corpus is structured by language, domain, and category, so licenses can be scoped to Arabic + Legal, Bahasa + Engineering, or any combination. Minimum license volumes apply. Contact us to discuss your requirements.

02How do you handle PII in healthcare and audio datasets?

Automated PII detection runs across all datasets before delivery. Audio uses speaker anonymisation. Healthcare records have patient identifiers removed at the source. Data lineage documentation is available for HIPAA, GDPR, and EU AI Act compliance review.

03Do you offer sample packs before full licensing?

Yes. 10% sample packs are available for textbooks (per language group) and coding. Audio samples are available for specific language-industry combinations. Healthcare samples are available under NDA with IRB-equivalent institutional review. Request them through the audit form.

04Can you build custom datasets we don't see here?

Yes. Custom corpus collection, annotation, and curation for domains or languages outside the standard catalogue. Minimum engagement sizes apply, and lead time depends on language, domain, and volume. Request an audit to scope custom engagements.

05What delivery formats are supported?

Audio: WAV/FLAC + JSON transcripts + RTTM diarisation + metadata CSV. Textbooks: JSONL (gzip) + Parquet metadata manifest, HuggingFace-compatible. Healthcare: DICOM + PDF + anonymised metadata. Coding: JSONL per language + problem taxonomy JSON. All formats work with Megatron-LM, NeMo, LLaMA-Factory, and HuggingFace datasets.

06How do you document data provenance for AI Act compliance?

Every textbook carries its ISBN. Every audio hour carries its enterprise data agreement reference, language code, industry vertical, and channel configuration. Every medical record carries its institutional source and modality. Data cards delivered with every dataset include provenance documentation in the EU AI Act Article 10 format.

EU AI Act Article 10 · Data provenance

11Contact

Tell us your data requirements

Tell us what you are building, which modalities you need, and where the model is weak. We reply with a corpus match, sample pack, or custom collection plan.

Response window
One business day
Email
hello@trainjust.com
Sample packs
Samples on request

Send the brief. We reply within a business day.

We reply within one business day.