Skip to content
Noise floor
Foundation · May 2026

Ralph-1, and a number for what “real improvement” means

Before a network can crown the best training recipe, it needs to know what counts as better at all. So the first real run wasn’t about the model — it was about measuring the noise.

01 The baseline

The baseline

Ralph-1
253.9M
params · GPT-2 BPE
Training
262M tokens
FineWeb-Edu · 69 min / 1×H100
Final loss
3.8163
bf16 validation

The canonical recipe trained Ralph-1 — 253,872,128 parameters, 262,144,000 FineWeb-Edu tokens (~262M, from a 1B-token tokenized corpus) — to a final validation loss of 3.8163 in bf16, in 69 minutes on a single H100. The model is the byproduct. What mattered was the calibration that came with it.

02 The noise floor

The noise floor

Seeds
10 × 500
steps, same config
Noise
σ = 0.0064
val-BPB, seed-to-seed
“Decisive” bar
2σ = 0.013
the king-change threshold

We trained the same 125M configuration across 10 seeds, 500 steps each, and measured the spread in held-out val-BPB. The standard deviation was σ = 0.0064. That sets the bar: a proposed recipe change has to beat the sitting king by more than 2σ = 0.013 val-BPB to count as signal rather than seed luck. That threshold isn’t a vibe — it lives in the validator code as a literal constant, and it’s what every king change since has had to clear.

03 Why measure noise before anything else

Why measure noise before anything else

You can’t call a recipe change an improvement until you know how much the score moves on seed alone. So before scoring a single proposed change, we pinned the seed-to-seed variation of the unchanged recipe. That number — 2σ = 0.013 val-BPB — is what turns ‘decisively beats the king’ from a vibe into a literal threshold in the validator code. Every king change since has had to clear it, and every evaluated change, win or loss, is published with its multi-seed variance. Honest limit: this is a single-box H100 measurement on FineWeb-Edu, and Ralph-1 is a deliberately short baseline — its job is to anchor the lineage cheaply, not to be a strong model.

Ralph-1 = 253,872,128 params · 262M FineWeb-Edu tokens · github.com/RalphLabsAI/ralph