Ethio-BBPE / training_metrics.json
Nexuss0781's picture
Production release: EthioBBPE tokenizer with perfect Amharic reconstruction
af2f42f verified
Raw
History Blame
1.34 kB
{
"model_info": {
"name": "EthioBBPE",
"version": "1.0.0",
"type": "Byte-level BPE",
"library": "tokenizers"
},
"training_data": {
"datasets": [
{
"name": "Synaxarium Dataset",
"source": "Nexuss0781/synaxarium",
"file": "synaxarium_dataset.parquet",
"num_texts": 366
},
{
"name": "Canon Biblical Amharic-English",
"source": "Nexuss0781/conon-biblical-am-en",
"file": "canon_biblical_am_en.parquet",
"num_texts": 61403
}
],
"total_files_processed": 2,
"combined_corpus_size_mb": 15.43,
"total_lines": 61769
},
"training_config": {
"vocab_size": 16000,
"min_frequency": 2,
"special_tokens": [
"<pad>",
"<unk>",
"<s>",
"</s>",
"<mask>"
],
"lowercase": false,
"compression": "gzip (level 9)",
"checkpointing": true
},
"performance_metrics": {
"perfect_reconstruction": true,
"test_samples": 3,
"amharic_geez_punctuation": "✅ Perfect",
"synaxarium_text": "✅ Perfect (66 tokens)",
"biblical_text": "✅ Perfect (82 tokens)",
"compression_ratio": "89.8% (1.3MB -> 136KB)"
},
"training_timestamp": "2026-05-01T23:41:22.748426",
"hardware": {
"cpu_cores": "auto-detected",
"training_time_seconds": 16.5
}
}