EXPERIMENT RECORD

Viperthon training history

The experiment record includes failed runs as well as recovered and current runs.

RUN A

Collapsed

13.4%

Final token accuracy. The embedding table collapsed while loss kept improving; the run finished below a word-count baseline, the Markov floor of 19.1%.

Final token accuracy 13.4%

RUN B

Recovered

43.7%

Same data and budget as Run A. The run recovered and reached 43.7% final accuracy, 3.26× Run A.

Full-context 44.95%

Shuffled-context control 3.61%

RUN B′

Current model

48.9%

Live accuracy for the current Viperthon Gen-1 B′ model. Training is at 344M of a 415M-token budget.

Live accuracy 48.9%

Tokens processed 344M / 415M

Vocabulary 36,001

Tokenizer Case-preserving

CONTROL METRICS

Run B

Control metrics recorded for the recovered experiment.

Metric	Value
acc_full	44.95%
acc_ctx1	2.84%
acc_shuf	3.61%
ctx_gain	42.11%
markov1 floor	19.09% (Run A)

TRAINING TRAJECTORY

Run B

Step 500

26.45%

Step 1,000

31.00%

Step 19,000

48.50%

Step 20,000

48.75%

Step 21,000

48.90%

RECORD NOTE

All numbers come from the training reports of the runs themselves — no re-measuring, no selection.