opengrad

Plugin opengrad from marketplace opengrad in arjhinety/OpenGrad. Commands read from the repo's manifest — yesterday.

Skills (11)

• opengrad-annotation

  How to run and change OpenGrad's annotation system safely — the opengrad-annotate CLI (check, start, status, audit, export, freeze-gold, model-batch, model-ingest, reference, review-queue, adjudicate, verify), the P-DET-v1 task, blinding of model judgments, the hash-chained SQLite store that only th

• opengrad-data-pipeline

  How to change OpenGrad's data pipeline correctly — source adapters and adapter versioning, the canonical IR and supervision contracts, the normalization-v3 builder (build, verify, determinism, audits, re-anchoring docs), canonical releases and mixtures, rendering and loss masks, and the opengrad-dat

• opengrad-development

  Entry point for any development work in the OpenGrad repository — repository map, environment (Windows, Git Bash, .venv, uv), how to verify a change the way CI does, known pre-existing failures, commit and push rules, and which opengrad-* skill owns each area. This skill should be used at the start

• opengrad-evaluation

  How to measure models correctly in OpenGrad — the immutable B0 baseline and the separate candidate path, the frozen held-out partitions, checkpoint selection rules frozen before results, evaluation methodology (response-mode coverage, answer_rate vs accuracy_given_answer, truncation, deterministic s

• opengrad-experiments-readiness

  How OpenGrad experiments are configured, gated and recorded — experiment configs, opengrad preflight, opengrad readiness and its gate sets (ready_for_baseline / ready_for_sft / ready_for_dpo), opengrad gpu-smoke, the authoritative run store (runs/, ExperimentStore, ledgers) versus the derived result

• opengrad-openweights-release

  How to publish and derive OpenGrad artifacts correctly — Hugging Face dataset and model releases with publication records and cards, the authoritative publication gate (scripts/verify_publication.py) and repository hygiene scan, GGUF conversion, quantization and parity, ExecuTorch exports, the optim

• opengrad-promotion-gates

  How to write, change and test gates in OpenGrad so they cannot pass vacuously — promotion policies (must_pass / max_regression / minimum_improvement, tool_use_promotion_v3/v4/v5/v6, quantization_preservation_v1), the Study 002 gate study_002_gate_v1, mandatory response-mode coverage and the anti-pat

• opengrad-registry-provenance

  How OpenGrad's registries, provenance claims and written records stay trustworthy — registry YAML and JSON schemas validated by opengrad-validate, the immutable-anchor rule for every verified claim, authoritative versions, generated documentation views, reports and hash-pinned artifacts, errata and

• opengrad-research-guardrails

  Research-integrity rules for OpenGrad — the G1–G17 guardrails, study freezing and errata, preregistration and amendments, frozen artifacts that must never change, phase gating by the study owner, and how to report numbers, negative results and incidents. This skill should be used before changing any

• opengrad-skills-maintenance

  How the OpenGrad development skills, shipped as the `opengrad` Claude Code plugin in plugins/opengrad/, stay correct as the repository grows — which skill owns which area, when a change must update a skill in the same commit, the tests/skills/test_skills.py coverage, reference and manifest checks th

• opengrad-training

  How to change or run OpenGrad training correctly — the real SFT and DPO loops and their fail-closed configuration contracts, the rendered-sample cache, resume and checkpoint lineage, checkpoint retention (upload before delete), the full-model component policy (vision + native MTP carried and trained

Similar plugins