colm-artifact-evaluation

COLM Artifact Evaluation

COLM had no formal artifact-evaluation track verifiable for the 2026 cycle (checked

2026-07-08; 待核实 each edition). That absence does not lower the bar — it moves the

audit into ordinary review, where artifact quality influences scores without a

rubric to appeal to. Package as if a skeptical reviewer will spend ten minutes with

your materials, because at this venue one usually will.

The COLM artifact taxonomy

LM papers produce artifact types with very different release mechanics; inventory

yours before deciding anything:

| Artifact | Review-time form | Release-time form | Blocking question |

|---|---|---|---|

| Code (training/eval) | Anonymized repo or supplement ZIP | Public repo, tagged release | Does one command reproduce one table? |

| Prompts | Verbatim appendix + files in package | Same, public | Exact strings, incl. system prompts? |

| Fine-tuned weights | Usually described, not uploaded (size) | Model hub upload with model card | Does the base model's license permit derivative release? |

| Training/eval data you built | Anonymized sample + datasheet | Full release with license | Any personal data, scraped ToS conflicts, or annotator-privacy issues? |

| Cached model outputs | Sample in supplement | Full archive | Does the provider's ToS permit publishing outputs at this scale? |

| Human-eval materials | Instructions + interface in appendix | Same | IRB/consent status stated? |

The two questions authors most often skip are in the right-hand column: derivative

weight licensing and output-publication ToS. Both can void a promised release after

acceptance — resolve them before the paper commits to anything.

One-command reproduction

The credibility core of the package is a single entry point per headline result:

# Makefile at package root — one target per main-text table/figure
table2:            ## headline comparison, ~40 GPU-min or ~$8 API spend
	python run_eval.py --config eval/run-042.yaml --out results/table2.csv
figure3:           ## scaling curve from cached outputs (no model access needed)
	python plots/scaling.py --cache outputs/cache.jsonl --out figs/figure3.pdf
verify:            ## regenerate all numbers from cached outputs only
	python verify_from_cache.py --tolerance 0.1

The verify-from-cache target is the LM-specific trick: reviewers without GPUs or

API budgets can still confirm that your published numbers follow from your recorded

model responses. It converts "trust me" into "check me" at zero compute cost, and it

keeps working after API models drift or deprecate (colm-reproducibility).

Anonymous review packaging

• Build from a clean export and grep for identity leaks — the commands and channel

list live in colm-supplementary; org-scoped model-hub IDs are the leak class

unique to LM work.

• If weights must be inspectable at review time, an anonymized hub account or a

size-reduced distilled checkpoint are the workable options; a link to your lab's

account is a double-blind violation under COLM's no-identifying-links rule.

• Include a MANIFEST.md: inventory, license per item, compute needed per target,

and what is deliberately absent with the reason ("training corpus omitted:

contains licensed text; filtering scripts included instead").

Datasheets for anything you release

For each dataset or evaluation set: how items were created (author population,

LLM-generated fraction — disclosable under the 2026 LLM policy), collection dates

(this doubles as contamination documentation for future users), license and source

licenses, known biases and coverage gaps, and a contamination canary if you want

future training runs to be detectable. For model releases: intended use, evaluation

scope, and known failure modes. These documents are cheap at packaging time and

impossible to reconstruct honestly later.

Post-acceptance release sequence

1. De-anonymize the repository; restore real hub org names and W&B links.
2. Upload weights/data with cards and licenses; tag the exact code release cited in

the camera-ready.

1. Add the paper's OpenReview URL to every artifact so provenance points both ways.
2. Freeze a colm2026 git tag — the paper cites a snapshot, not a moving main.
3. Update the paper's availability statement to match what actually shipped

(colm-camera-ready owns the deadline: August 7, 2026 this cycle).

The ten-minute reviewer walkthrough

Dry-run the package as the busiest plausible reviewer before every upload. The

sequence they follow is predictable, so optimize for it in order:

1. Minute 1-2: open MANIFEST.md. If there is no manifest, they grep for a

README and form their opinion from whatever half-stale file they find. The

manifest is the cheapest score you will ever buy.

1. Minute 3-4: look for the entry point. make table2 or an equivalent single

command. If the first thing they see is a 400-line setup guide with cluster

assumptions, the walkthrough ends here.

1. Minute 5-7: run verify from cache. No GPUs, no keys, under a minute of

compute — this is the step that actually gets executed in practice, which is

why the cached-outputs archive earns its place in the package.

1. Minute 8-9: spot-check a prompt file against the paper's appendix. Any

mismatch between packaged prompts and printed prompts contaminates trust in

both.

1. Minute 10: skim the code for identity leaks and hardcoded secrets — partly

ethics-duty, partly curiosity. This is where an org-scoped from_pretrained

string ends your anonymity.

Sizing note: keep the review package lean — cached outputs can be sampled down to

what verify needs, with the full archive promised for release. No supplementary

size cap was verifiable for COLM 2026 (待核实), but a multi-gigabyte upload fails

socially even where it succeeds technically.

Longevity: the two-year test

A COLM artifact's real audience arrives later — the group in two years trying to

compare against you. Two cheap investments serve them: freeze an environment

manifest (exact package versions; a container digest if you can), and write the

MANIFEST.md assuming every external URL in it will eventually rot — name

artifacts by content hash where possible so mirrors stay verifiable. The venue

publishes on OpenReview, where your artifact links are permanently attached to the

paper's public page; links that die quietly are the failure mode, so prefer

archival hosts for anything you would want cited.

Output format

[Inventory] code ▢ prompts ▢ weights ▢ data ▢ cached-outputs ▢ human-eval ▢
[Legal gates] base-model license: <ok/blocks release>  output-ToS: <ok/limits>
[One-command check] targets exist for: <tables/figures>  verify-from-cache: ▢
[Anonymity] package clean / leaks: <items>
[Manifest + datasheets] present / missing: <which>
[Release plan] <ordered steps with dates>