brycewang-stanford COLM Reproducibility
Reproducibility at a language-modeling venue is a different problem than at a classical ML venue, because the central artifact — the model — is often either too large to rerun or owned by someone else and changing weekly. COLM papers therefore need a layered story: what any reader can redo, what a funded lab can redo, and what will drift no matter what anyone does. Say which layer each claim lives in.
The three reproducibility layers
| Layer | Who can redo it | Your obligation |
|---|---|---|
| Exact | Anyone with the artifact | Pinned checkpoint + harness + prompts + decoding → same numbers (within stated sampling noise) |
| Budgeted | A lab with comparable compute | Full training recipe: data, hyperparameters, schedule, hardware, wall-clock, known failure modes |
| Drifting | Nobody, eventually | API-model results: version string, query dates, and an expiry-honest caveat in the paper |
A paper whose headline claim lives entirely in the drifting layer needs to say so on the page where the claim is made — and ideally to anchor the finding on at least one open-weight model so the exact layer is not empty.
API models: the honesty protocol
Closed models are legitimate scientific objects at COLM — much of the field studies them — but they demand a specific discipline:
- Record the full version identifier and the date range of every query; "GPT-class model, spring 2026" is not a citation.
- Treat model deprecation as a known risk: state which results a reader can expect to reproduce only while the version is served.
- Do not average over silent version changes — if a provider updated mid-experiment, the runs before and after are different systems.
- Cache raw responses. The response archive is the reproducible artifact when the model disappears; releasing (input, output) pairs, licenses permitting, converts a drifting claim into an exact one about the recorded behavior.
What "pinned" means, concretely
# An open-weight result is pinned when all four hashes exist in the paper/repo:
MODEL_REV=8f3c21a # checkpoint revision, not just the model name
TOKENIZER_REV=8f3c21a # tokenizers change under stable names too
HARNESS_SHA=$(git -C eval-harness rev-parse --short HEAD)
PROMPTS_SHA=$(sha256sum prompts/*.txt | sha256sum | cut -c1-12)
echo "model=$MODEL_REV tok=$TOKENIZER_REV harness=$HARNESS_SHA prompts=$PROMPTS_SHA"
The harness hash matters more than newcomers expect: extraction rules, answer normalization, and few-shot formatting differ across harness releases and can move benchmark scores by points. A score without a harness version is not comparable to anyone else's score.
Training-side recipe (budgeted layer)
If the paper trains or fine-tunes, report: base checkpoint and revision; dataset composition with filtering steps and, where possible, released data or reproducible filtering code; optimizer, learning-rate schedule, batch size, steps/epochs, precision; hardware and GPU-hours; and everything that went wrong — divergences, restarts, hand-tuned interventions. The failure narrative is what separates a recipe someone can follow from a recipe that only worked once.
Compute and cost disclosure
Give a single table: training GPU-hours (or API fine-tuning spend), evaluation GPU-hours or API token spend, and totals including preliminary experiments if material. This is a COLM-community expectation for two reasons — cost is part of any efficiency claim, and disclosure lets smaller groups judge whether replication is even feasible before they try.
The availability statement
Write it as commitments, not vibes. For each artifact class — code, prompts, trained weights, training data, evaluation data, raw model outputs — state one of: released at submission (anonymized), released on acceptance, available on request, or cannot be released because <specific reason: license, privacy, ToS>. "Code will be made available" with no object and no date is the pattern reviewers have learned to distrust.
Note: no dedicated COLM reproducibility checklist or artifact-certification track was verifiable for the 2026 cycle (checked 2026-07-08, 待核实 each edition) — this discipline is enforced by reviewers, not by a form, which in practice makes it more visible in scores, not less.
When you cannot release
Legitimate blockers exist; the discipline is to substitute the strongest releasable proxy instead of going silent:
| Blocked artifact | Reason | Releasable proxy |
|---|---|---|
| Training corpus | Licensed / scraped text | Filtering code + corpus statistics + a datasheet of sources |
| Fine-tuned weights | Base-model license forbids derivatives | Training recipe + LoRA/adapter deltas if permitted + eval harness |
| Raw API outputs | Provider ToS limits republication | Scored per-item results + a small quoted-excerpts sample within fair use |
| Human-eval transcripts | Annotator privacy | Aggregated ratings + instructions + anonymized examples with consent |
| Internal eval set | Planned reuse as a held-out set | A public dev split + hash commitments to the held-out items |
The hash-commitment trick in the last row deserves wider use: publishing item hashes today proves later that the held-out set did not move, which protects both you and the benchmark.
Pre-submission audit
- Every number's layer identified (exact / budgeted / drifting), and the paper says which.
- Four hashes present for every open-weight result; version strings + query dates for every API result; raw responses cached.
- Training recipe complete enough that its absence would be the reviewer's only question.
- Compute table present.
- Availability statement itemized by artifact class with concrete commitments.
- Anonymization of released artifacts checked (
colm-submissionsweep) — an identity-leaking repo is worse than none at submission time.
Output format
[Repro layers] exact: <claims> / budgeted: <claims> / drifting: <claims>
[Pinning audit] model ▢ tokenizer ▢ harness ▢ prompts ▢ query-dates ▢
[Response cache] exists / missing for API experiments
[Compute table] present / absent
[Availability] itemized commitments / vague — rewrite
[Top gap] <the one omission a reviewer will find first>