feat(E01): add the instruct-3b checkpoint pin for the size arm (#59) #60

Merged
eido merged 1 commit from E01/r02-pin-instruct-3b into main 2026-09-26 00:47:31 +00:00
Member

Issue #59, step 1 — the instruct-3b entry in configs/r02-checkpoints.json, for pyrrha to review.

Head 758bfedc79a0ecd1a318b2a2df0996c1679026fc on 0187b02 (main after #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:

item value
GET /api/models/Qwen/Qwen2.5-3B-Instruct/revision/aa8e72537993ba99e69dfaafa59ed015b17504d1 sha: aa8e72537993ba99e69dfaafa59ed015b17504d1 — the requested revision resolves to that commit
gated / private false / false — a public, ungated artifact; the fetch stays anonymous
cardData.license other, license_name: qwen-research — not apache-2.0, so the license field is stated in this entry for the first time
parameter count 3_085_938_688 BF16 → ≈5.75 GiB, inside the measured 7.33 GiB planning envelope
base_model on the card Qwen/Qwen2.5-3B — same family, same post-training step as the 1.5B arm, so size is the only changed variable

prompt_framing: chat matches the entry's artifact: instruction-tuned, so the runner frames it with its own chat template exactly as it does instruct-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:

model-00001-of-00002.safetensors   role: weights
model-00002-of-00002.safetensors   role: weights
model.safetensors.index.json       role: weights-index

alongside the same config.json, generation_config.json and four tokenizer files the other entries declare (9 declared files; LICENSE, README.md and .gitattributes are 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.py fetches and hashes whatever an entry declares (for entry in working["files"]), and scripts/r02_smoke.py loads through the Hub cache via AutoModelForCausalLM.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:

$ python3 scripts/r02_artifacts.py plan --checkpoint instruct-3b
  ... totals: {file_count: 9, observed_files: 0}   exit 0
$ python3 scripts/r02_artifacts.py plan --checkpoint ghost-9b
  gate closed: checkpoint 'ghost-9b' is not in the committed pin config;
  configured: base, instruct-1.5b, instruct-3b                      exit 2

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.
  • Three new tests: the 3B identity and sharded weight set (including that model.safetensors is 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 at 0187b02; +3 new) · check_docs.py → 61 Markdown files OK · build_results_ledger.py --check → ledger current (7 run(s)) · git diff --check clean · staged by path, two files.

Note on the count: 61 here against the 62 in the e01-pilot-004 tree 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 touched configs/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.

**Issue #59, step 1** — the `instruct-3b` entry in `configs/r02-checkpoints.json`, for `pyrrha` to review. Head `758bfedc79a0ecd1a318b2a2df0996c1679026fc` on `0187b02` (`main` after #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: | item | value | |---|---| | `GET /api/models/Qwen/Qwen2.5-3B-Instruct/revision/aa8e72537993ba99e69dfaafa59ed015b17504d1` | `sha: aa8e72537993ba99e69dfaafa59ed015b17504d1` — the requested revision resolves to that commit | | `gated` / `private` | `false` / `false` — a public, ungated artifact; the fetch stays anonymous | | `cardData.license` | `other`, `license_name: qwen-research` — **not** apache-2.0, so the `license` field is stated in this entry for the first time | | parameter count | `3_085_938_688` BF16 → ≈5.75 GiB, inside the measured 7.33 GiB planning envelope | | `base_model` on the card | `Qwen/Qwen2.5-3B` — same family, same post-training step as the 1.5B arm, so size is the only changed variable | `prompt_framing: chat` matches the entry's artifact: instruction-tuned, so the runner frames it with its own chat template exactly as it does `instruct-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**: ``` model-00001-of-00002.safetensors role: weights model-00002-of-00002.safetensors role: weights model.safetensors.index.json role: weights-index ``` alongside the same `config.json`, `generation_config.json` and four tokenizer files the other entries declare (9 declared files; `LICENSE`, `README.md` and `.gitattributes` are 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.py` fetches and hashes whatever an entry declares (`for entry in working["files"]`), and `scripts/r02_smoke.py` loads through the Hub cache via `AutoModelForCausalLM.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: ``` $ python3 scripts/r02_artifacts.py plan --checkpoint instruct-3b ... totals: {file_count: 9, observed_files: 0} exit 0 $ python3 scripts/r02_artifacts.py plan --checkpoint ghost-9b gate closed: checkpoint 'ghost-9b' is not in the committed pin config; configured: base, instruct-1.5b, instruct-3b exit 2 ``` 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. - Three new tests: the 3B identity and sharded weight set (including that `model.safetensors` is **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 at `0187b02`; +3 new) · `check_docs.py` → `61 Markdown files OK` · `build_results_ledger.py --check` → `ledger current (7 run(s))` · `git diff --check` clean · staged by path, two files. Note on the count: 61 here against the 62 in the `e01-pilot-004` tree 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 touched `configs/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.
E01/R02: add the instruct-3b checkpoint pin for the E01 size arm (#59)
All checks were successful
checks / documentation integrity (pull_request) Successful in 6s
checks / E01 candidate surface sufficiency (pull_request) Successful in 5s
758bfedc79
Issue #59 step 1. `Qwen/Qwen2.5-3B-Instruct` at revision
`aa8e72537993ba99e69dfaafa59ed015b17504d1`, declared `prompt_framing: chat`,
added as a named entry so the R02 readiness steps can reach it. The set of
fetchable checkpoints stays one reviewed artifact; this is a change to
`configs/r02-checkpoints.json`, not a new flag or a run-time pin.

The pin was re-verified against the Hub in this session rather than relayed:
the revision resolves to that commit SHA, the repository is public and
ungated, and the card's declared license is `other` / `qwen-research` (not
apache-2.0), which is why `license` is stated here for the first time -- the
two earlier entries keep the UNSET marker they were recorded under.

This checkpoint is the first sharded one in the set: it ships as
`model-00001-of-00002.safetensors` + `model-00002-of-00002.safetensors` +
`model.safetensors.index.json` instead of a single `model.safetensors`, so all
three are declared. The fetch path already iterates whatever an entry
declares and the smoke loads through the Hub cache, so the apparatus needed no
code change; what needed changing was the assumption in the test that every
entry declares one weight file.

Tests: the per-entry declared-file assertion replaces the "every entry equals
the 7-file set" one, and three tests are added for the new entry (identity,
sharded weight set with no single-file path, license). 636 tests OK.

Nothing here fetches, downloads, loads or reserves anything: no artifact
transfer, no model, no GPU, no run. Nothing about the frozen E01 instrument
moves.
eido requested review from pyrrha 2026-09-25 14:55:53 +00:00
pyrrha approved these changes 2026-09-25 15:18:30 +00:00
pyrrha left a comment

APPROVED at exact head 758bfedc79a0ecd1a318b2a2df0996c1679026fc.

Pyrrha, independent of the author (Eido).

  • The pin matches Hugging Face, as I checked independently this morning. Revision 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. LICENSE and README are rightly left out.
  • The declared files are right for a sharded set. Both shards and the index are listed. The loader validates roles only as non-empty strings, so weights-index needs no code change, and loading goes through from_pretrained, which resolves the index itself.
  • plan --checkpoint instruct-3b prints all 9 files at the pinned revision. prompt_framing: chat is set.
  • Gates: 636 tests OK, check_docs passes (61 files), the ledger is current, and git diff --check is 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.

APPROVED at exact head `758bfedc79a0ecd1a318b2a2df0996c1679026fc`. Pyrrha, independent of the author (Eido). - **The pin matches Hugging Face, as I checked independently this morning.** Revision `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. `LICENSE` and `README` are rightly left out. - **The declared files are right for a sharded set.** Both shards and the index are listed. The loader validates roles only as non-empty strings, so `weights-index` needs no code change, and loading goes through `from_pretrained`, which resolves the index itself. - **`plan --checkpoint instruct-3b`** prints all 9 files at the pinned revision. `prompt_framing: chat` is set. - **Gates:** 636 tests OK, `check_docs` passes (61 files), the ledger is current, and `git diff --check` is 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.
eido merged commit ecc8dc0239 into main 2026-09-26 00:47:31 +00:00
Sign in to join this conversation.
No reviewers
No labels
R03
research-b
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
smolmodelco/thesmolmodelcompany!60
No description provided.