Real human datafor agenticLLM training.
Human Data by ForwardLane is a privacy-controlled training dataset for agentic large language models.
We ship ten production-grade slices drawn from a real five-year operational graph — joined, canonicalised, pseudonymised, provenance-back-linked, and sealed with a three-axis privacy budget — to AI labs and applied-AI teams under an NDA. Read our privacy policy and terms of service.
An operational dataset, not a benchmark.
Human Data v0.1 is a single, signed pack of ten production-grade slices extracted from a real five-year operational graph at ForwardLane — the same graph that runs the company. Every entity is canonicalised, every edge is bi-temporal, every passage carries a PROV-O back-link to its source URI, and the whole pack ships with three orthogonal privacy gauges so labs can audit before they train.
What it is. A single tar of Parquet + ZSTD files (with a JSONL mirror for ergonomic streaming), one slice per task — joined across four source layers: an operational CRM/PM graph, a long-document store, a code+ontology corpus, and a stream of OpenTelemetry agent traces.
How it was built. Live data is pulled from operational systems, entity-resolved with deterministic blocking rules, canonicalised against a 15-ontology reference (BFO · FIBO · UFO · PROV-O · SDLC · CWE · OWASP and friends), HMAC-pseudonymised against a per-slice key, then sliced along ten task axes. No synthetic infill, no scraping.
Why bi-temporal. Every edge carries four timestamps — t_valid, t_invalid, t_created, t_expired. Contradictions invalidate prior facts rather than overwrite them, so an agent can reason about what was true at a given moment, not just what is true now.
Distribution. Each release is a signed bundle: SHA-256 over the bytes, Ed25519 over the manifest, fingerprint published. A reviewer can audit any slice without re-running the pipeline.
Access is NDA-gated. The pack is provided under the Human Data License — sub-licensing, model-publication, and retention terms are covered in the agreement.
- Pack
- forwardlane_v1
- Version
- v0.1.0 — pilot
- Slices
- 10 task-aligned
- Coverage
- ~5 years operational history
- Sources
- 14 enterprise systems
- Entities
- ~39K canonical (contacts + companies + deals + tasks)
- Documents
- 15,341 (~38 GB markdown)
- Code
- 13.6M LOC across 151 repos
- Edges
- 70,019 ontology mappings · 15 ontologies
- Traces
- 22,915 OTel spans · 6,698 unique IDs
- Bi-temporal
- t_valid · t_invalid · t_created · t_expired
- Format
- Parquet + ZSTD · JSONL mirror
- Size
- ~40 GB compressed · ~120 GB JSONL
- Hashes
- SHA-256 per file
- Signature
- Ed25519 (operator fingerprint published)
- Privacy
- G·01 · G·02 · G·03 · G·04
- License
- Human Data License — NDA gated
Operational, not client.
Human Data is sourced exclusively from ForwardLane's own operational graph — the systems that run our company. No client data of any kind is included in any slice, in any release, ever.
ForwardLane does not have access to, store, process, or derive training data from client data in any form. Client information stays in the client's environment under their controls. The pack contains zero rows, fields, or derivatives that originated from a customer dataset.
All client traffic crosses TLS 1.3; all client data at rest sits behind AES-256 envelope encryption with per-tenant keys. Decryption requires the client's key material — ForwardLane operators cannot read plaintext.
The pack is ForwardLane's own operational record. Before any byte enters a slice, every quasi-identifier is replaced with an HMAC-derived pseudonym keyed per release. Canonical URIs never leave our resolver.
Every release passes a multi-screen privacy budget: k-anonymity ≥ 5 and ℓ-diversity ≥ 3 on sensitive attributes, Wasserstein-1 leakage banded per slice, LiRA membership-inference at FPR=10⁻³, and Carlini canary extraction bounded. No PII, no reconstruction, no extraction with current state-of-the-art methods.
This boundary is a load-bearing contract, not a marketing claim. The gauges below are the receipts.
What rides inside the pack.
Every entity carries a JSON-LD @type array. Every passage carries a PROV-O back-link. Every edge carries (t_valid, t_invalid, t_created, t_expired) — contradictions invalidate rather than overwrite.
25.5K contacts · 6.2K companies · 87 deals across HubSpot, Salesforce, Asana and Jira — joined into one canonical knowledge graph with HMAC-keyed pseudonyms.
~38 GB · ~4.5B markdown characters from Drive and iCloud. 7,326 organisations and 4,044 persons extracted with provenance back to source URIs.
70,019 mapping edges across 15 ontologies — BFO, FIBO, UFO, PROV-O, SDLC, CWE, OWASP — on 110 repos. 43% of latest-revision functions carry ≥ 1 mapping.
6,698 unique trace IDs · 21,117 embedded passages. Every passage carries a PROV-O back-link to its source document URI.
Three orthogonal gauges, one signed manifest.
A reviewer audits any slice without re-running the pipeline: recompute SHA-256 over the bytes, confirm the gauge band matches the value, verify the operator's ed25519 signature against the published fingerprint.
Per-slice keys rotated per release. Canonical URIs never leave the resolver — only HMAC-derived pseudonym IDs ship in the pack.
Each slice carries a band — tight / moderate / wide — for distributional distance from the held-out reference. Villani 2009; scipy bootstrap CI.
Likelihood-ratio attack at FPR = 10⁻³ against a shadow model. We ship the band, not the raw score. Carlini et al. 2022.
Secret-Sharer canaries planted at controlled rarity; we publish the 95th-percentile exposure as an upper bound. Carlini 2019 + 2021.
What you actually get.
Each release ships the same load-bearing capabilities. The research that backs each one is cited inline — every claim is auditable, no hand-waving.
Reason about what was true then, not just now. Every edge carries (t_valid, t_invalid, t_created, t_expired); contradictions invalidate rather than overwrite — agents can train on the history of a fact, not just its latest state.
Cross-system identity stitched once, traceable forever. 25.5K contacts across HubSpot · Salesforce · Asana · Jira collapsed to 14K canonical entities with an audit trail per merge — no manual re-review.
One embedding, three resolutions (64 / 256 / 1024 dim). Cheap retrieval for filtering, full-fat for re-ranking, same vector — pick the latency-quality point per query.
Every passage and every fact carries a PROV-O link back to its source URI. Reviewers (and your future agents) can trace any claim to the document, line, and timestamp that produced it.
Every release passes a red-team pass: LiRA membership inference at FPR=10⁻³ (gauge G·03) and Carlini Secret-Sharer canary extraction (gauge G·04). We ship the bands, not the raw scores.
SHA-256 per file, Ed25519 over the manifest, operator fingerprint published. A reviewer can audit any slice without re-running the pipeline — just recompute, confirm, verify.
Want the underlying papers and standards in full? See the deck appendix — the foundations are public; only the pipeline that joins them is proprietary.
We turn operational reality into training signal.
ForwardLane is an applied-AI company. For years we've built and run production AI inside real enterprises — knowledge graphs, agents, and decision systems operating against live operational data, not benchmarks. We know what messy, multi-source, privacy-bound business data actually looks like, because we've shipped against it.
Now we package that experience for the labs. We help AI labs and companies training agents with provenance-grade, privacy-controlled datasets — and the tooling to evaluate agentic reasoning on data that behaves like the world their agents will actually operate in: bi-temporal, contradictory, entity-resolved, and back-linked to source.
Real human data, joined and canonicalised by people who've done this in production — so your agents train on reality, not a sandbox.
Sourced from live operational systems and unified through a deterministic, reproducible joiner — not scraped, not synthetic-by-default.
HMAC pseudonymisation, k-anonymity / ℓ-diversity, and three orthogonal leakage gauges shipped in every signed manifest.
A team that has deployed enterprise AI in regulated, real-world environments — that judgement is encoded in every slice.
Want the full pack?
Request access and we'll send you the Human Data License & NDA to sign. Access opens upon execution.
Request access