BDH-25M-PL training baseline

Verify that the official BDH architecture trains end-to-end on a Polish byte corpus and establish an inspectable 25M-scale baseline.

01 · Design

Method before metrics.

Train for 10,000 steps on 100M cleaned Polish byte tokens and publish the checkpoint, configuration and training outcome.

Question: RQ-0001 · Which evaluations provide useful signal for 25M–500M language models?
Hypothesis: H-0001 · Validation bits-per-byte changes smoothly before downstream exact-match benchmarks separate 25M-class models.

02 · Protocol

A reproduction matches this protocol. A fork changes it and declares the deviation.

model Family: Dragon Hatchling (BDH)
parameter Count: 25,300,000
train Tokens: 100,000,000
dataset: SlayerLab/polish-dynaword-mix
controlled Variables: official BDH architecture · byte vocabulary · sequence length
independent Variable: training step
seeds: 42
benchmarks: validation byte-level cross entropy

03 · Evidence outputs

Contribute

Start locally, inspect the portable manifest, and sign in only if you choose to publish.

Add evidence

Record the measurement and its provenance. New evidence is unverified and pending review until CodeSOTA checks it.

04 · Record