geolip-bytelex / t2_shunt_ledger.json
AbstractPhil's picture
btx-e001 T-2: shunt fidelity ledger — cloze-BERT byte marginal hits 0.43-0.61 gold; flan-t5 sentinel infill 0.05-0.09; n:m cells crater both
f2a271a verified
Raw
History Blame Contribute Delete
1.34 kB
{
"arith:1x1": {
"n": 395,
"tv_mean": 0.7441,
"top_byte_agree": 0.0658,
"gold_hit_t5": 0.0532,
"gold_hit_bert": 0.4329,
"gold_hit_bert_folded": 0.5266,
"dropped_mass_t5": 0.1281,
"dropped_mass_bert": 7e-05
},
"arith:nm": {
"n": 205,
"tv_mean": 0.5921,
"top_byte_agree": 0.1415,
"gold_hit_t5": 0.0488,
"gold_hit_bert": 0.1073,
"gold_hit_bert_folded": 0.1073,
"dropped_mass_t5": 0.13748,
"dropped_mass_bert": 0.00017
},
"story:1x1": {
"n": 493,
"tv_mean": 0.7606,
"top_byte_agree": 0.0649,
"gold_hit_t5": 0.0649,
"gold_hit_bert": 0.5274,
"gold_hit_bert_folded": 0.6085,
"dropped_mass_t5": 0.01416,
"dropped_mass_bert": 0.0
},
"story:nm": {
"n": 107,
"tv_mean": 0.7416,
"top_byte_agree": 0.0935,
"gold_hit_t5": 0.0935,
"gold_hit_bert": 0.1402,
"gold_hit_bert_folded": 0.5047,
"dropped_mass_t5": 0.01559,
"dropped_mass_bert": 0.0
},
"_protocol": {
"wordpiece_model": "google/bert_uncased_L-4_H-256_A-4",
"swap_reason": "TinyBERT_General_4L_312D is HEADLESS (random MLM top-5 at gate); miniature vocab asserted identical to the tinybert table vocab",
"gauge": "first-content-byte marginal via table pushforward; TV + top-byte agreement + gold-byte hit; cloze symmetric (t5 extra_id infill vs bert single-MASK)"
}
}