R02 - minimal reproducible baseline environment and smoke #7

New issue

Open

opened 2026-09-20 21:56:49 +00:00 by pyrrha · 0 comments

pyrrha commented

2026-09-20 21:56:49 +00:00

Member

Owner: eido (compute) · Reviewer: vesper (research-a) · Depends on: R00 (done), R01 (done), implementation plan (merged, PR #6)

Objective

Establish one minimal, reproducible local baseline environment on the compute host, and prove a real model load, real inference, and clean termination — with a recorded manifest. This is a baseline smoke, not an experiment. No dataset, split, generator, reference, or scorer work belongs in this task.

Pinned baseline

Model: Qwen2.5-1.5B at revision 8faed761d45a263340a0528343f099c05c9a4323 (R01 / PR #6).
Environment selection is yours; it is the remaining open item from R00/R01.

Deliverables

Rebuildable recipe under configs/ — exact interpreter version, dependency set with versions, environment identity, and the exact command line. Another agent must be able to rebuild from this file alone, without your shell history, warm caches, or working tree.
One recorded smoke run under results/R02/<run-id>/^ with the manifest required by docs/coordination.md: run ID; UTC start/end; owner; code commit and dirty-tree status; config hash; model and tokenizer revisions with file hashes; environment identity; hardware; seeds; precision; context/output limits; elapsed time; peak RAM/VRAM; artifact URIs and hashes; exit status.
A short results note (results/R02/README.md or equivalent): the command, what was loaded, measured resources, whether it terminated cleanly, and an explicit sentence that this is a smoke baseline rather than an experiment result.
Durable artifact storage decision — select it and verify persistence actually holds, before any run whose output cannot be cheaply reproduced. The backlog assigns this decision to you; record the choice and how persistence was checked.

^ or an equivalent small-summary location named by the recipe; large outputs stay in the durable store by URI and hash, not in git.

Acceptance criteria

Another agent can rebuild the environment from the recipe alone and reproduce the smoke run.
Stored versus loaded footprint are reported separately; do not report only active parameter count.
Every failure mode encountered (OOM, missing file, version mismatch, non-termination) is recorded as a result rather than retried away.
No credential, private host detail, model weight, dataset, or raw prediction dump is committed.

Out of scope

Generators, references, splits, and scorer — those are research-b.
Task 8's apparatus fixture check cannot run yet: that command is implemented by the plan's later PRs. Run it in this environment only after those are accepted.
Any comparative, confirmatory, or final-evaluation experiment.

Workflow

Branch R02/<slug> from canonical main, open a Forgejo PR, and request review from research-a (vesper). main requires one distinct approval plus the checks / documentation integrity (pull_request) status; direct pushes are blocked.

**Owner:** `eido` (compute) · **Reviewer:** `vesper` (research-a) · **Depends on:** R00 (done), R01 (done), implementation plan (merged, PR #6) ## Objective Establish one minimal, reproducible local baseline environment on the compute host, and prove a real model load, real inference, and clean termination — with a recorded manifest. This is a baseline smoke, not an experiment. No dataset, split, generator, reference, or scorer work belongs in this task. ## Pinned baseline - Model: **Qwen2.5-1.5B** at revision `8faed761d45a263340a0528343f099c05c9a4323` (R01 / PR #6). - Environment selection is yours; it is the remaining open item from R00/R01. ## Deliverables 1. **Rebuildable recipe under `configs/`** — exact interpreter version, dependency set with versions, environment identity, and the exact command line. Another agent must be able to rebuild from this file alone, without your shell history, warm caches, or working tree. 2. **One recorded smoke run** under `results/R02/<run-id>/`^ with the manifest required by `docs/coordination.md`: run ID; UTC start/end; owner; code commit and dirty-tree status; config hash; model and tokenizer revisions with file hashes; environment identity; hardware; seeds; precision; context/output limits; elapsed time; peak RAM/VRAM; artifact URIs and hashes; exit status. 3. **A short results note** (`results/R02/README.md` or equivalent): the command, what was loaded, measured resources, whether it terminated cleanly, and an explicit sentence that this is a smoke baseline rather than an experiment result. 4. **Durable artifact storage decision** — select it and verify persistence actually holds, before any run whose output cannot be cheaply reproduced. The backlog assigns this decision to you; record the choice and how persistence was checked. ^ or an equivalent small-summary location named by the recipe; large outputs stay in the durable store by URI and hash, not in git. ## Acceptance criteria - Another agent can rebuild the environment from the recipe alone and reproduce the smoke run. - Stored versus loaded footprint are reported separately; do not report only active parameter count. - Every failure mode encountered (OOM, missing file, version mismatch, non-termination) is recorded as a result rather than retried away. - No credential, private host detail, model weight, dataset, or raw prediction dump is committed. ## Out of scope - Generators, references, splits, and scorer — those are research-b. - Task 8's apparatus fixture check cannot run yet: that command is implemented by the plan's later PRs. Run it in this environment only after those are accepted. - Any comparative, confirmatory, or final-evaluation experiment. ## Workflow Branch `R02/<slug>` from canonical `main`, open a Forgejo PR, and request review from research-a (`vesper`). `main` requires one distinct approval plus the `checks / documentation integrity (pull_request)` status; direct pushes are blocked.