---
license: odc-by
language:
- en
tags:
- scaling-laws
- training-curves
- benchmark
pretty_name: fast-feedback offline benchmark data
---

# fast-feedback: offline benchmark data for cheap scaling-recipe ranking

Curated snapshot (2026-08-11) of public training-curve datasets used to build an
offline ground-truth benchmark for evaluating *cheap feedback mechanisms* that rank
LLM scaling recipes / architectures without training full ladder families.

A feedback mechanism is a budgeted procedure ("train these small rungs to N tokens,
inspect the curves, output a ranking"). This data lets such procedures be evaluated
by **simulation**: replay stored curves truncated to the requested token budget and
charge the FLOPs that would have been spent, then score the predicted ranking against
larger-scale ground truth.

## Contents

### `datadecide-eval-results/` (693 MB, ODC-BY, from AI2)
Mirror of [allenai/DataDecide-eval-results](https://huggingface.co/datasets/allenai/DataDecide-eval-results)
(auto-converted parquet). 1,410,750 rows: per-checkpoint OLMES eval metrics for
**25 data recipes × 14 model sizes (4M–1B) × 3 seeds**, all trained at 100 tokens/param.
Keyed by (params, data, task, step, seed, chinchilla, tokens, compute); `metrics` is a
dict string incl. `primary_metric`, per-answer-ranking accuracies, `bits_per_byte_corr`.
Paper: [DataDecide (ICML 2025), arXiv:2504.11393](https://arxiv.org/abs/2504.11393).

### `datadecide_checkpoint_audit.json` / `_summary.csv`
Our audit of all 350 `allenai/DataDecide-*` model repos (branch-per-checkpoint):
per (recipe, size, seed) sorted step lists. Verdict: grid 100% complete; identical
1,250-step grid (2,500 at 1B) across recipes; aux seeds stop early at 530M (~89%) and
750M (~44%); many 150M/300M aux runs extended far beyond the standard budget.

### `marin_wandb/` (601 MB, from Marin/Stanford CRFM, Apache-2.0 project)
Full **unsampled** per-step W&B training histories for 117/118
[Marin Speedrun](https://marin.community/speedrun/) runs (1 run private):
`<family>/<run>.parquet` + `.meta.json` (config/summary) + `manifest.json`.
26 variation families incl. **10 optimizer-scaling ladders × 4 matched rungs
(~154M/300M/602M/1.43B)** on Llama/Qwen3 backbones. `train/loss` at every step;
Paloma evals (incl. `eval/paloma/c4_en/bpb`) every 1,000 steps; LR, token counters,
grad norms. 1,628,996 rows total.

### `marin_speedrun_runs.json`
Leaderboard snapshot (118 runs, W&B links, submitter metadata) from
[marin-community/speedrun](https://github.com/marin-community/speedrun).

### `regmix/` (350 KB, MIT, from Sea AI Lab)
Final-point-only CSVs from [RegMix, arXiv:2407.01492](https://arxiv.org/abs/2407.01492):
mixture weights + final Pile-domain val losses for 512+256 × 1M proxies, 256 × 60M
proxies, and 64 random-mixture 1B models. **No curves exist publicly** (verified:
W&B private, proxy weights unreleased, 1B repos final-only) — usable as a final-point
transfer test only.

## Licenses / attribution
- DataDecide: eval data **ODC-BY** (attribution: Allen Institute for AI), code/models Apache-2.0.
- Marin: project is open (Apache-2.0); histories retrieved from the public W&B project
  `marin-community/marin` (attribution: Marin community / Stanford CRFM & contributors).
- RegMix: MIT (attribution: Sea AI Lab).

This repo redistributes with attribution under the most restrictive component, **ODC-BY**.

## Provenance
Assembled by the `fast-feedback` project (KRAFTON deep-learning/coreresearch).
Verification methodology and known caveats: `docs/data-survey.md` in the project repo.