nanochat English vs Greek medium experiment

This repo stores the 353M nanochat English vs Greek comparison. The English baseline is retained from the March 25, 2026 run, and the Greek side was refreshed on March 30, 2026 with dedup replay plus mojibake filtering using fffoivos/glossapi-greek-nanochat-pretraining-dataset.

Final loss summary:

  • English final validation bpb: 0.890841
  • English held-out test bpb: 0.891791
  • Greek final validation bpb: 0.300441
  • Greek held-out test bpb: 0.331570

Validation BPB Comparison

Validation bpb comparison

Training Loss Comparison

Training loss comparison

  • Delta (greek - english) final validation bpb: -0.590400
  • Delta (greek - english) held-out test bpb: -0.560221

Artifacts:

  • models/english/final/
  • models/greek_threshold_050/final/
  • tokenizers/english/
  • tokenizers/greek_threshold_050/
  • analysis/
  • docs/RUN_REPORT.md
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support