README.md

metadata

license: gpl-2.0
language:
  - en
tags:
  - scaling_laws
  - pre_training
  - llms
  - language_modeling
  - llama
paper: >-
  Tokens-per-Parameter Coverage Is Critical for Robust LLM Scaling Law
  Extrapolation
venue: NeurIPS 2026
configs:
  - config_name: default
    data_files:
      - split: train
        path: '**/*.json'
dataset_info:
  features:
    - name: checkpoint_prefix
      dtype: string
    - name: timestamp
      dtype: string
    - name: status
      dtype: string
    - name: model_size
      dtype: int64
    - name: d_model
      dtype: int64
    - name: n_layers
      dtype: int64
    - name: n_heads
      dtype: int64
    - name: d_ff
      dtype: int64
    - name: vocab_size
      dtype: int64
    - name: max_seq_len
      dtype: int64
    - name: learning_rate
      dtype: float64
    - name: epochs
      dtype: int64
    - name: batch_size
      dtype: int64
    - name: weight_decay
      dtype: float64
    - name: seed
      dtype: int64
    - name: deterministic
      dtype: bool
    - name: cudnn_deterministic
      dtype: bool
    - name: cudnn_benchmark
      dtype: bool
    - name: total_tokens
      dtype: int64
    - name: total_steps
      dtype: int64
    - name: cumulative_time_seconds
      dtype: float64
    - name: val_epochs_count
      dtype: int64
    - name: mean_val_loss
      dtype: float64
    - name: best_val_loss
      dtype: float64
    - name: final_train_loss
      dtype: float64
    - name: std_train_loss
      dtype: float64
    - name: mean_train_loss
      dtype: float64
    - name: global_batch_tokens
      dtype: int64
    - name: final_val_loss
      dtype: float64

Collinear/Non-Collinear Scaling Models

Checkpoint repository for scaling law experiments comparing collinear (CO) and non-collinear (NC) experimental designs for the paper Tokens-per-Parameter Coverage Is Critical for Robust LLM Scaling Law Extrapolation under review for NeurIPS 2026.

Code

Anonymized code repository (reproduces all tables): anonymous.4open.science

Directory Structure

{dataset}/{design}/N_{param_count}/

• Dataset: wikipedia, pes2o, cosmopedia, redpajama, c4 (plus _bf16 and _bigtpp variants)
• Design: colinear or non_colinear
• N: Model parameter count (one of 14 canonical sizes from ~5M to ~76.5M)

Experimental Designs

Collinear (CO): Models are trained along a line in (N, D) space where D = TPP × N for varying TPP (tokens per parameter) values. A single model size N is swept across many TPP values.

Non-collinear (NC): Models are trained on a grid over (N, D) space (NxD_GRID), varying both N and D independently.

Holdout Sets

Some checkpoints include HOLDOUT in the filename. These were held out from scaling law fitting and are used to evaluate extrapolation / interpolation accuracy of fitted scaling laws. Both CO and NC designs have holdout checkpoints:

• COLINEAR_HOLDOUT_* → collinear holdout (held-out TPP values)
• *_HOLDOUT_* (without COLINEAR) → non-collinear holdout (held-out (N, D) pairs)

Filename Convention

{PREFIX}{DESIGN}N{approx_size}[TPP{val}]D{tokens}_{dataset}_m{exact_N}_token{exact_D}lr{lr}..._completedAt{timestamp}.pt

The m{N} and token{D} fields contain the exact parameter count and token count used for training.

Full results pre-extracted are available in this repository aswell, under: extracted_losses/ , full_coverage_results/, prelim_bigtpp_results/, seed_variance_results/, special_results/ and subset_bbox_seeds/ .

---

License: GPL v2.0