ajaxdavis/alpha-er-corpus

alpha-er training corpus

The exact 3.86 GB text mixture used to train ajaxdavis/alpha-er, a 100M-parameter model trained end to end on a from-scratch, CUDA-free GPU stack.

This is published as the precise artifact that was trained on, byte for byte, so the run is reproducible rather than approximately describable.

Composition

A single <|end_of_text|>-delimited plain-text file, interleaved from three sources in proportional blocks so that every source is represented throughout — including the validation tail, which a naive concatenation would have left containing only one source.

Source	Bytes	Share	Upstream	License
FineWeb-Edu / DCLM / FinePDFs	1,992,298,579	51.7%	HuggingFaceFW/finepdfs_edu_50BT-dclm_30BT-fineweb_edu_20BT-shuffled	ODC-By 1.0
SmolTalk	1,163,711,296	30.2%	HuggingFaceTB/smol-smoltalk	Apache-2.0
Concordance-EN	699,311,635	18.1%	dense vocabulary-coverage English prose (ours)	CC-BY-SA-4.0
Total	3,855,321,510

The pretraining half is the first four retained parquet shards of the FineWeb mixture, 1,857,705 accepted documents (17,508 skipped), built 2026-07-22. The pretrain-000 shard used here has SHA-256 d993342b0bb55198c520f1f761bb0aad2812b2d8fb9c6347b4e6f9d622794d9c.

Licensing

CC-BY-SA-4.0, and the reason is worth stating plainly: Concordance-EN is share-alike, so the mixture inherits its terms even though the other two components are more permissive. Attribute all three upstreams listed above.

Chat format

SmolTalk contributes the conversational structure, using three special tokens:

<|user|>What is the capital of France?<|assistant|>Paris.<|end_of_text|>

Use in training

alpha-er consumed 1.97B tokens from this corpus — roughly 0.9 epochs at a 12,288-entry byte-level BPE vocabulary — in 20,000 steps of 98,304 tokens, reaching validation loss 4.48 (perplexity ~88).

The resulting model writes fluent English in the right register but is not factually reliable. Anyone reproducing it should expect the same: the corpus is large enough to teach syntax and register, and a 100M model with 320 active FFN units per token is not large enough to retain much of its content.