fffoivos/nanochat-en-vs-el-353m

nanochat English vs Greek medium experiment

This repo stores the 353M nanochat comparison runs. It now includes the original English baseline, the original Greek run, and a refreshed cleaned Greek run produced with the same nanochat recipe and builder path as the original Greek comparison.

Three-way Results

Run	Final validation BPB	Best validation BPB	Train BPB	Test BPB	Best step	Train time
English baseline	0.890841	0.890841	0.852060	0.891791	2708	2.49 h
Original Greek	0.300441	0.300441	0.307879	0.331570	2708	0.87 h
Refreshed cleaned Greek	0.424307	0.424307	0.401086	0.434514	2708	0.93 h

Delta for refreshed cleaned Greek versus original Greek:

• Final validation BPB: +0.123866
• Best validation BPB: +0.123866
• Train BPB: +0.093207
• Test BPB: +0.102944

Three-way Plots

Refreshed Greek Data

• Dataset repo: fffoivos/glossapi-greek-nanochat-pretraining-dataset
• Dataset revision used for this run: e3b8d19d40665551d49af369efb082e82b6a815e
• Builder script: /home/ubuntu/experiments/nanochat/2026-05-07_refreshed-dataset-comparable-353m/runs/greek-refreshed-dataset-comparable-d13/build/scripts/prepare_glossapi_greek_experiment_data_v2.py
• Train/validation/test char targets: 5658225610 / 251485449 / 251485449
• Split namespace and seed: nanochat-en-vs-el-medium, seed 20260322
• Quality filters: greek_badness_score < 10.0 and mojibake_badness_score <= 0.1
• Dedup replay: drop_intra_and_inter, exact stage strict_and_relaxed, near threshold 0.85, policy share_aware
• Markdown chunking and shuffled chunks are enabled, matching the original Greek recipe.

Artifacts

• models/english/final/
• models/greek_threshold_050/final/
• models/greek_refreshed_dataset_comparable/final/
• tokenizers/english/
• tokenizers/greek_threshold_050/
• tokenizers/greek_refreshed_dataset_comparable/
• analysis/loss_comparison_refreshed_dataset_three_way.json
• analysis/final_manifest_refreshed_dataset_comparable.json
• docs/RUN_REPORT.md
• docs/RUN_REPORT_REFRESHED_DATASET_COMPARABLE.md