artifacts/standard-named

Candidate Public Training Profile

artifacts/standard-named/20260715__TELIC-FIELDS__PROFILE__WORKING__F-11-G-12__public-training-profile.md

Rendered from markdown source. Open raw source on GitHub.

Candidate Public Training Profile

Status: research artifact Content canon status: unset

A public profile should disclose enough to evaluate training claims without exposing sensitive source data or security details.

Identity and release

model_family:
model_version:
provider:
release_regime:
intended_purposes:
prohibited_purposes:

Source authority coverage

Report percentages or bounded qualitative coverage for:

SOURCE KNOWN
LICENSE KNOWN
AUTHORITY KNOWN
COMMUNITY GOVERNANCE REVIEWED
MATERIAL UNKNOWN
MATERIAL CONTESTED

Source-class table

For each material class:

source_class:
provenance_status:
authority_basis:
purpose_scope:
license_or_governance:
privacy_status:
withdrawal_support:
benefit_terms:
known_disputes:

Human contribution

Disclose:

  • contributor roles;
  • recruitment region and structure;
  • compensation method;
  • psychological-risk controls;
  • disagreement handling;
  • downstream-use notice.

Preference and constitution

Disclose:

  • RLHF, DPO, RLAIF, or other method;
  • evaluator population and coverage;
  • aggregation rule;
  • disagreement handling;
  • constitution authorship;
  • revision authority;
  • protected standing;
  • known gaps.

Synthetic data

Disclose:

  • generating models;
  • approximate proportion;
  • source lineage;
  • recursive depth;
  • contamination and diversity testing.

Withdrawal and unlearning

Disclose supported rungs:

COLLECTION STOP
SOURCE DELETION
FUTURE EXCLUSION
RUNTIME RETRIEVAL BLOCK
APPROXIMATE UNLEARNING
RETRAINING
MODEL RETIREMENT

State verification limits.

Succession and derivatives

Disclose:

  • derivative models;
  • open-weight implications;
  • restrictions and duties propagated;
  • responsible successor.

Benefit and accountability

Disclose:

  • contributor and community benefit;
  • audit;
  • challenge;
  • repair;
  • responsible contact.

Material caveat

This profile describes training lineage. It does not establish that every deployment of the model is consentful, lawful, safe, or legitimate.