TPPIsCritical's picture
Normalised run tables for the public-checkpoint reanalysis
592c6db verified
|
Raw
History Blame Contribute Delete
2.07 kB
metadata
license: cc-by-4.0
task_categories:
  - text-generation
tags:
  - scaling-laws
  - language-models
  - experimental-design

Public-checkpoint reanalysis: normalised run tables

Normalised training-loss observations from published language-model studies, used for the public-checkpoint reanalysis accompanying Tokens-per-Parameter Coverage Is Critical for Robust LLM Scaling Law Extrapolation.

Every number in that analysis is recomputable from these tables alone, with no need to re-fetch the upstream sources.

Files

File Rows Contents
runs_expanded.csv 12,531 Full corpus: all ten qualifying studies, including mid-training checkpoints
runs_normalised.csv 753 Terminal-observation corpus used for the primary analysis
datadecide_ppl_by_group.csv - DataDecide per-group perplexities with reconstructed token counts
farseer_1222_full.csv - Farseer released run table

Schema

cell (study/corpus identifier), run_id, N (non-embedding parameters), D (training tokens), loss, loss_smooth, loss_val where released, and k = D/N (tokens per parameter).

Provenance

Each row derives from the releasing study's own published loss data. Model size and token count are recoverable per run; studies where either could not be recovered, or which span too few distinct model sizes to carve a design from, were screened out and are logged with the requirement they failed in the accompanying analysis package.

Included studies: Farseer, Chinchilla (isoFLOP, digitized), Gemstones (annealed and constant-LR), Gadre et al. over-training, the OLMo ladder, DataDecide, MAD, GPT-3 curves, and OPT trajectories.

The ColPret aggregation (Choshen et al.) is a source for several of these; it is not redistributed here, and the fetch script in the analysis package pulls it from the original release.

Reproducing

The analysis code, the pre-registered protocol with its dated amendments, and the screening log accompany the paper. Point the analysis at these tables in place of the fetch step.