The behavioural ground truth
for commercial AI.

Generic LLM training data doesn't teach commercial behaviour.

Foundation models are trained on internet text and synthetic role-plays. Commercial conversations are something else entirely: diagnosing problems, framing value, handling objections, building rapport under pressure, recovering from a missed signal. None of that is reliably present in the corpora today's models learn from.

Built to the standard that commercial AI actually needs.

English mirror on every quote. Original language preserved.

Commercial conversations don't only happen in English; most training corpora pretend they do. Every primitive carries the quote in its source language beside an English mirror, so you can fine-tune for one market without losing the transferability of the English label.

// One turn. Non-English calls carry quote_original + ISO code.
"rep_behaviour": {
  "speaker_role":         "the rep",
  "language_original":    "en",
  "quote_original":       null,
  "quote_english":        "Before I answer, one check: is the concern
                       the setup work itself, or when it lands?",
  "observable_behaviour": "The rep paused to disambiguate the buyer's
                       constraint before answering, treating effort
                       and timing as separate objections."
}

A single skill vocabulary across five commercial functions.

Peer v2 is a multi-SEA ontology (Skill–Evaluation Architecture) covering every customer-facing commercial role. Each function has its own competencies and stages; all share one tier ladder, one evidence structure and one scoring contract. That is what makes a single dataset useful across the whole commercial-AI surface.

From a graded call to structured signal.

Below: a representative primitive in the exact 2026.1 export shape — structure, fields and grading exactly as the production exporter emits them; the conversation itself is fictionalised, because real corpus records are licensed, not published. Every primitive carries a verification object and is operator-verified before it enters the licensable corpus. Every primitive follows this shape, including next_tier_example, the same shape a licensee receives.

// One real licensable primitive, taken verbatim from the downloadable sample. Long strings truncated for display only.
{
  "primitive_id": "prim_2026-05-18_e21c07ba_5f43d2-042",
  "schema_version": "1.2",
  "sanitisation_version": "v1_sanitise-v0.1",
  "ontology_version": "peer_v2",
  "sanitised_at": "2026-05-19T09:41:12.104583+00:00",
  "function": "SALES",
  "framework_code": "AE-MIDMRKT",
  "framework_display_name": "Mid-Market AE Sales Framework",
  "competency_code": "S-SEA CC9",
  "competency_display_name": "Listening and Responsiveness",
  "cla_code": "LI",
  "skill_theme": "Listening & Responsiveness",
  "function_tags": [],
  "grade": "PROFICIENT",
  "polarity": "POSITIVE",
  "exchange": {
    "rep_behaviour": {
      "speaker_role": "the rep",
      "quote_english": "Before I answer, one check: is the concern the setup work itself, or when it lands? Those pull in different directions.",
      "quote_original": null,
      "language_original": "en",
      "observable_behaviour": "The rep paused before answering to disambiguate the buyer's constraint, recognising that setup effort and go-live timing are two different objections wi…"
    },
    "prospect_context": {
      "speaker_role": "the buyer",
      "quote_english": "we'd want this in place before the new team starts, but the diary is the real problem. Onboarding lands the same month, so whatever you propose has to f…",
      "quote_original": null,
      "language_original": "en"
    },
    "prospect_reaction": {
      "speaker_role": "the buyer",
      "quote_english": "It's the timing. If it lands after week two, the setup work is fine.",
      "quote_original": null,
      "language_original": "en"
    }
  },
  "verification": {
    "status": "verified",
    "verified_at": "2026-06-02T10:17:44.612905+00:00"
  },
  "licensable": true,
  "license_terms": null,
  "rubric_evidence": "The rep listened closely enough to spot the ambiguity in the buyer's constraint and disambiguated before answering, a clear Proficient-level listening b…",
  "next_tier_example": {
    "quote": "Just so I answer the right thing: is it the setup effort, the timing, or both? And I think I'm also hearing that the new team's first week matters more …",
    "explanation": "Disambiguates the surface question and also names the underlying concern (the new team's first impression) the buyer hadn't yet articulated, which is t…"
  }
}

# Load with HuggingFace datasets
from datasets import load_dataset

ds = load_dataset("json",
                  data_files="peerlibrary-sample.jsonl")

# Inspect a primitive
print(ds["train"][0]["competency_display_name"])
# → "Multi-Threading Strategy"

# Filter to ADVANCED-tier sales primitives
adv = ds["train"].filter(
    lambda r: r["ai_grade"]["tier_grade"] == "ADVANCED"
              and r["function"] == "SALES"
)

Train, fine-tune, evaluate on real ground truth.

Buy-vs-build, head to head.

What an ML team weighs PeerLibrary against: synthetic data, or an in-house labelling project on Scale or Labelbox. Both are real options. Both lose on the dimensions below.

Where every primitive comes from, and what's stripped before it ships.

The contract that makes the dataset legally distributable. Six steps from source recording to licensable primitive. Every primitive carries the sanitisation version that processed it, so licensees can re-pull under newer contracts.

Try first. Scale on a usage shape that makes sense.

We onboard partners selectively; the dataset gets better when the people training on it are good at their craft. Three tiers; pricing is bespoke and depends on volume, refresh cadence and integration depth.

Request a custom evaluation sample.

The 15-primitive sample shows the shape. Want the slice that matches your use case, by function, framework, grade mix or language? Tell us what you're evaluating and we'll build a private evaluation portal for your team: dataset, schema and datasheet. We verify each requester first, so it isn't instant. It is free, and tailored to you.

For the teams building the next layer of commercial AI.

PeerLibrary is licensed selectively. We onboard partners we can stand behind, and whose use cases improve the data for everyone.