docs(model-ports): proposed draft validation contracts for pending identities (issue #200) - #202
Conversation
…identities Stage 1 of docs/MODEL_PORTS_TRACK_PLAN.md needs a reviewed host oracle/ validation contract per new identity before it can be registered -- drafting these to save the review effort of writing 20 from scratch, not to self-certify anything. Placed under .github/ci/reference/proposed_model_ports/ (a new subdirectory, not the active .github/ci/reference/ contracts) so there's no ambiguity that these are proposals pending maintainer review, not already-trusted fixtures. Every field is either verified fact (upstream repo + 40-char pinned commit, license, the exact GGUF sha256 we downloaded and hashed locally) or explicitly marked measurement_status: pending_host_or_board_run with a note explaining why -- no PPL numbers or correctness thresholds are fabricated. We don't have board access to produce those; a real host/board run should populate them, not an assumed number. Covers the same 20 models as issue aifoundry-org#200 plus the follow-up comment. mpt7b is excluded here too -- same reason as its missing claim file (mosaicml/mpt-7b returns 401 anonymously, no pinnable revision).
…nite_3_1b_a400m, phi3_mini, minicpm3_4b Same draft-only pattern as the original 20 (see PR body). Unlike those, these 5 carry real measured host-CPU perplexity (measurement_status: measured_host_cpu_only) instead of a pending placeholder -- produced this session with a plain CPU-only (GGML_ET=OFF) llama-perplexity build against the board-pinned WikiText-2 corpus. ET-SoC1 board perplexity is still not measured; that requires the maintainer's own trusted workflow.
…4b, apertus_8b, glm4_9b (real measured PPL, see PR aifoundry-org#206)
…eal measured PPL, see PR aifoundry-org#207)
|
These proposed contracts need to be tied to the exact qualifying PR/artifact for each execution family, deduplicated against overlapping claims, and backed by reproducible committed oracle outputs/hashes and justified thresholds. Please remove entries whose corresponding port is incomplete or unclaimed. This proposal alone does not grant credit; maintainers must review and promote accepted contracts into the authoritative protected reference/identity ledger. |
…nclaimed proposed contracts Hackathon deadline has passed; this is record-completeness cleanup only. Per maintainer review, removed bloom560m and pythia410m -- neither has completed hosting or a registered board benchmark, so their proposed contracts don't correspond to a valid claim yet.
|
Note: this update was made after the official hackathon deadline and is not a hackathon submission or seeking credit — cleaning up for aiNekko's ET-SoC1 card records. Removed For the remaining 31 proposed contracts here: most of the corresponding model ports now have a real |
2 participants
Draft validation contracts for 20 model-port identity candidates (see #200). New files only, own subdirectory, not merged into active contracts.