---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
language:
- en
tags:
- marin
- delphi
- scaling-laws
- pretrained
- research-only
datasets:
- nvidia/Nemotron-CC
- bigcode/starcoderdata
- EleutherAI/proof-pile-2
---

# delphi-9e18-358Mparams-4.7Btokens

A 358M-parameter base model from the **Delphi** scaling suite. Trained at 9 × 10¹⁸ FLOPs on 4.7B tokens with the Delphi recipe.

## About Delphi

Delphi is the Marin team's first open scaling suite, inspired by
[Pythia](https://www.eleuther.ai/papers-blog/pythia-a-suite-for-analyzing-large-language-modelsacross-training-and-scaling).
It has three parts:

- a **scaling recipe** that maps compute budgets to model configurations,
- a **scaling suite** of models trained from that recipe at IsoFLOP budgets
  from 3 × 10¹⁸ to 1 × 10²³ FLOPs, and
- a **scaling law** which uses the smaller Delphi models to predict the larger
  ones.

A pre-registered forecast from that scaling law predicted the final loss of the
largest Delphi run (1 × 10²³ FLOPs, 25 B parameters, 600 B tokens) within
**0.2%**, using **300× less compute** than the training run itself. The same
process forecasts downstream benchmarks — MMLU, HumanEval, and GSM8K — via a
two-step regression combining compute and observational scaling laws.

See ["Scaling Laws That Extrapolate 300× Past the Fit"](https://openathena.ai/blog/delphi)
for the recipe, fit, and downstream-eval projections. The full set of Delphi
checkpoints — IsoFLOP grid points, held-out optima at 1e21/1e22/1e23 with
multiple random seeds, and training intermediates — lives on
[`marin-community`](https://huggingface.co/marin-community) on the Hub.

This is a research artifact, not a production model.

## Model details

| | |
|---|---|
| Architecture | Qwen 3 (pre-norm decoder, RMSNorm, RoPE, QK-norm with learned scaling, SwiGLU MLPs) |
| Parameters | 358,306,688 |
| Hidden size | 896 |
| Layers | 10 |
| Attention heads | 7 |
| KV heads | 7 (no GQA) |
| Head dim | 128 |
| FFN intermediate | 3584 (MLP ratio 4) |
| Vocab size | 128,256 (Llama 3 tokenizer) |
| Max sequence length | 4096 |
| Position encoding | RoPE (θ = 500000, Llama 3-style scaling) |
| Bias terms | None |
| Tied embeddings | No |

## Training

| | |
|---|---|
| Compute | 9 × 10¹⁸ FLOPs |
| Tokens | 4,728,946,688 |
| Steps | 36,078 |
| Sequence length | 4096 |
| Optimizer | AdamH (Adam with Hyperball) |
| Recipe | Delphi (Complete(d)P-style scaling with `(T₀/T)^0.3` token-horizon LR correction) |
| LR schedule | WSD: 10% linear warmup, 20% linear decay, 0 floor |
| Precision | f32 master params, bf16 compute |
| Parallelism | FSDP |
| Data mixture | Nemotron-CC + StarCoderData + ProofPile 2 |
| Tokenizer | Llama 3 (vocab 128,256) |

[AdamH](https://whenwen.github.io/wd_blog/public/hyperball-part-1.html), Adam
with Hyperball, constrains every projection weight to stay on the Frobenius-
norm sphere it was initialized on, so weight decay has nothing to regularize
away and falls out of the recipe. A Complete(d)P-style transfer rule with a
`(T₀/T)^0.3` correction sets learning rate as token horizon grows. Reference
constants: B₀ = 64, T₀ = 2.5 B tokens, η₀ = 0.00630, η₀,Adam = 0.000656,
ε₀ = 1.85 × 10⁻⁸. Recipe code:
[`experiments/scaling_law_sweeps/completed_adamh.py`](https://github.com/marin-community/marin/blob/main/experiments/scaling_law_sweeps/completed_adamh.py).

## Companion releases

- All Delphi model checkpoints: [`marin-community`](https://huggingface.co/marin-community) on the Hub.
- Plot data behind every figure in the blog post:
  [`marin-community/delphi-blog-data`](https://huggingface.co/datasets/marin-community/delphi-blog-data) (one config per figure, with `wandb_url` on every row).
- Pipelines that deterministically reproduce the training mixture from public
  Nemotron-CC, StarCoderData, and ProofPile 2 sources: see the Marin repo.

## Evaluation

This checkpoint is part of the Delphi eval suite
([`experiments/exp1337_eval_suite.py`](https://github.com/marin-community/marin/blob/main/experiments/exp1337_eval_suite.py)),
which scores every Delphi run alongside reference open-weights baselines
(Qwen 3, Llama 2/3, OLMo 2, Marin 8B). Following the blog's two-step
forecast, soft metrics (per-choice log-prob for multiple-choice tasks,
bits-per-byte for generative tasks) carry the signal the scaling law is fit on,
and a sigmoid fit on an external model pool maps soft metric to hard metric
(accuracy, pass@1, exact-match). Below ~1e21 FLOPs the hard metrics stay near
chance even when the underlying probabilities are improving smoothly; that is
expected and is exactly why the soft metrics exist.

## Limitations

- Trained on an English-heavy web mixture; no multilingual coverage.
- Pretrained-only — no instruction tuning, RLHF, or safety alignment.
- The Delphi recipe targets compute-optimal training, not inference-cost-aware
  overtraining; for inference-heavy deployments, an overtrained smaller model
  may be preferable. The blog's "off-optimal training" section quantifies the
  penalty.
- This is one checkpoint in a much larger Delphi release; pick the one that
  matches your compute / parameter / token regime, or browse the full set at
  [`marin-community`](https://huggingface.co/marin-community).

## Citation

```bibtex
@misc{held2026delphi,
  title  = {Scaling Laws That Extrapolate 300× Past the Fit},
  author = {Held, Will and {Marin Community}},
  year   = {2026},
  url    = {https://openathena.ai/blog/delphi}
}
```