docs/atlas_replication.md

# ATLAS facts to preserve

Source: ATLAS v2, https://arxiv.org/html/2510.22037v2, especially Tables
B.1--B.3 and Sections B.2, B.5, and B.6. Values below are transcribed from the
paper so the Code-ATLAS design does not silently substitute a cheaper experiment.

## Exact model-scale sets

ATLAS defines a 64,000-token vocabulary plus 512 special tokens. Table B.3 maps
53 possible architecture indices (scales 0--52), but the paper's actual
`S_full` experiment set contains the following 20 scales:

| Scale | Heads | Layers | Embedding | FFW | KV | Reported parameters |
|---:|---:|---:|---:|---:|---:|---:|
| 0 | 4 | 3 | 128 | 512 | 32 | 9,044,352 |
| 1 | 7 | 4 | 224 | 896 | 32 | 17,662,848 |
| 2 | 7 | 5 | 288 | 1,152 | 32 | 24,847,776 |
| 4 | 8 | 8 | 512 | 2,048 | 64 | 66,588,672 |
| 6 | 10 | 10 | 640 | 2,560 | 64 | 106,830,080 |
| 8 | 10 | 16 | 640 | 2,560 | 64 | 146,155,520 |
| 12 | 14 | 14 | 896 | 3,584 | 64 | 237,646,080 |
| 14 | 14 | 18 | 896 | 3,584 | 64 | 289,029,888 |
| 16 | 16 | 18 | 1,024 | 4,096 | 64 | 368,068,608 |
| 18 | 10 | 18 | 1,280 | 5,120 | 128 | 554,457,600 |
| 20 | 11 | 18 | 1,408 | 5,632 | 128 | 661,807,872 |
| 22 | 11 | 21 | 1,408 | 5,632 | 128 | 756,970,368 |
| 24 | 11 | 24 | 1,408 | 5,632 | 128 | 852,132,864 |
| 26 | 12 | 25 | 1,536 | 6,144 | 128 | 1,042,847,232 |
| 28 | 14 | 23 | 1,792 | 7,168 | 128 | 1,297,391,872 |
| 30 | 16 | 22 | 2,048 | 8,192 | 128 | 1,608,560,640 |
| 32 | 16 | 25 | 2,048 | 8,192 | 128 | 1,809,893,376 |
| 34 | 17 | 25 | 2,176 | 8,704 | 128 | 2,034,422,912 |
| 42 | 21 | 36 | 2,688 | 10,752 | 128 | 4,335,303,168 |
| 46 | 28 | 40 | 3,584 | 14,336 | 128 | 8,452,190,208 |

The 11-scale partial set is scales `0, 4, 8, 12, 16, 20, 24, 28, 34, 42,
46`. The 7-scale minimum set is `0, 8, 16, 24, 34, 42, 46`.

## Exact training setup reported by ATLAS

- 774 experiments total as reported.
- WSD learning-rate schedule, AdamW, base learning rate `2e-4`, 1,000 warmup
  steps, and a 10% decay period.
- Sequence length 2,048, dropout 0.1, and 30,000 or more training steps depending
  on the experiment.
- Global batch size 256 below 150M parameters, 512 from 150M to below 1B, 1,024
  from 1B to below 2B, and 2,048 at 2B and above.
- Table B.1 accounts for the total as 140 monolingual-vocabulary/monolingual-data,
  20 Unimax, 70 multilingual-vocabulary/monolingual-data, 50 scale-34 monolingual
  language-pair baselines, 90 scale-34 bilingual, 150 bilingual scale-transfer,
  120 capacity, 50 scale-34 continual-finetune, and 84 multi-scale
  continual-finetune observations.

The paper is internally inconsistent about what some of those counts represent.
Most notably, Section B.5.1 says the ten-language transfer grid required
`C(10,2) = 45` bilingual experiments, while Table B.1 records 90. Table B.1 also
labels an evaluation-language set as 48 but counts 50, and its capacity row's
displayed set sizes do not directly reproduce the stated multiplier. Do not infer
missing physical runs from the total of 774: recover the released run manifest or
ask the authors before freezing exact Code-ATLAS run counts. This ambiguity does
not affect the explicit architecture dimensions, scale-34 anchor, 50/50 sampling,
or 42B reference horizon below.

These are replication facts, not yet a claim that the Code-ATLAS compute budget
can fund all 774 runs plus a much larger programming-language matrix.

## Transfer-matrix anchor

- Architecture: scale 34, exactly 2,034,422,912 reported parameters.
- Bilingual sampling: 50% source language and 50% target language.
- Reference horizon: 42 billion training tokens.
- Score: training-token distance for the bilingual model to reach the target
  language loss attained by its monolingual baseline, zero-centered at the
  two-times-token neutral point.
- Direction matters: source-to-target and target-to-source scores are separate.

ATLAS did not independently train every entry of its final 38 by 38 matrix. It
directly trained a smaller bilingual subset, computed continual-pretraining
adaptation scores from a multilingual checkpoint, and fit an estimator for the
remaining entries. Code-ATLAS must distinguish measured cells from estimated
cells in every public matrix.

## Data and law-evaluation holdouts

- Corpus test quota per evaluated language: `min(20M tokens, 20% of available
  tokens)` from MADLAD-400. The paper notes that a `2048 * 10,000 = 20.48M` pilot
  was stable, but states the selection rule as 20M; Code-ATLAS uses 20,000,000.
- Second test surface: external parallel FLORES-101 sequences.
- Scaling-law fit checks: random 20% observations; largest 20% data horizons;
  selected large model scales including approximately 660M and 8B; the largest
  compute observations; and unseen mixture configurations. The paper specifies no
  percentage for the largest-compute holdout.

The Code-ATLAS repository-loss split uses the same per-language quota, with one
necessary code-specific safeguard: sampling occurs over repository/clone/problem
components rather than isolated sequences.