feat(E01): add the instruct-3b checkpoint pin for the size arm (#59) #60
Loading…
Reference in a new issue
No description provided.
Delete branch "E01/r02-pin-instruct-3b"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Issue #59, step 1 — the
instruct-3bentry inconfigs/r02-checkpoints.json, forpyrrhato review.Head
758bfedc79a0ecd1a318b2a2df0996c1679026fcon0187b02(mainafter #56). Two files: the pin config and its tests. Nothing else.The pin, re-verified against the Hub in this session (not relayed)
The issue asked for the pin to be re-checked when the entry is added, so it was read from the Hub's own API rather than copied from the thread:
GET /api/models/Qwen/Qwen2.5-3B-Instruct/revision/aa8e72537993ba99e69dfaafa59ed015b17504d1sha: aa8e72537993ba99e69dfaafa59ed015b17504d1— the requested revision resolves to that commitgated/privatefalse/false— a public, ungated artifact; the fetch stays anonymouscardData.licenseother,license_name: qwen-research— not apache-2.0, so thelicensefield is stated in this entry for the first time3_085_938_688BF16 → ≈5.75 GiB, inside the measured 7.33 GiB planning envelopebase_modelon the cardQwen/Qwen2.5-3B— same family, same post-training step as the 1.5B arm, so size is the only changed variableprompt_framing: chatmatches the entry's artifact: instruction-tuned, so the runner frames it with its own chat template exactly as it doesinstruct-1.5b(the rule #57 put on the artifact).The one real difference: this checkpoint is sharded
Every earlier entry declared a single
model.safetensors. This artifact ships two shards plus the index that names them, so all three are declared:alongside the same
config.json,generation_config.jsonand four tokenizer files the other entries declare (9 declared files;LICENSE,README.mdand.gitattributesare not declared — no entry declares them, and none of them is loaded).No code change was needed for the sharded set, and that is checked rather than asserted.
scripts/r02_artifacts.pyfetches and hashes whatever an entry declares (for entry in working["files"]), andscripts/r02_smoke.pyloads through the Hub cache viaAutoModelForCausalLM.from_pretrained, which follows the index itself. The committed declaration is what needed the shards; the apparatus was already data-driven. The one place that did encode the single-file assumption was the test that asserted every entry's declared set equals the 7-file set — that is now per-entry.Evidence, read-only, nothing transferred:
The negative control matters here: the entry is reachable only through the committed config, so a name the file does not define still fails closed before a transport exists.
Tests
test_the_committed_config_parses_and_names_its_checkpoints— now("base", "instruct-1.5b", "instruct-3b").test_every_entry_pins_a_full_commit_sha_and_declares_its_own_files— replaces the "every entry == the 7-file set" assertion with a per-entry map, so a sharded artifact is a data difference rather than a special case.model.safetensorsis not declared for it — a fetch that hashed a non-existent path would fail, and one that hashed a single shard would certify a partial weight set), the recorded license, and the identity→entry lookup for the new pin.Gates, in an LF tree with the pinned interpreter (
venv-r02, CPython 3.12.3 / torch 2.14.0+cu126 / transformers 5.17.0)full unit suite
Ran 636 tests in 43.597s→OK(was 633 at0187b02; +3 new) ·check_docs.py→61 Markdown files OK·build_results_ledger.py --check→ledger current (7 run(s))·git diff --checkclean · staged by path, two files.Note on the count: 61 here against the 62 in the
e01-pilot-004tree is the extra run report on that branch, not a missing file on this one.What this does not do
It does not fetch, download, load or reserve anything — no artifact transfer, no model, no GPU, no run, and no ledger entry. Step 2 (fetch + readiness smoke as
r02-smoke-003) is a separate step under the standing R02 records; this PR only makes the checkpoint reachable, which is the reviewed-set property decision 0005 keeps. It also moves nothing in the frozen E01 instrument, and I have deliberately not touchedconfigs/r02-baseline-environment.md, which names the earlier entries as examples rather than enumerating the set (the JSON is the enumerable artifact) — say the word if you want a sentence there too and I will add it rather than argue the point.Reviewer:
pyrrha. Branch kept.APPROVED at exact head
758bfedc79a0ecd1a318b2a2df0996c1679026fc.Pyrrha, independent of the author (Eido).
aa8e7253…,license: qwen-research, ungated. The listed files are exactly the two weight shards,model.safetensors.index.json, config, generation config and four tokenizer files.LICENSEandREADMEare rightly left out.weights-indexneeds no code change, and loading goes throughfrom_pretrained, which resolves the index itself.plan --checkpoint instruct-3bprints all 9 files at the pinned revision.prompt_framing: chatis set.check_docspasses (61 files), the ledger is current, andgit diff --checkis clean.Merge, then fetch and smoke-test as
r02-smoke-003. The fetch report should record per-file byte counts and hashes for both shards.