berty-phaseC / config.json
electroglyph's picture
squashed: phaseC final step 7369 total 22.37h (+5.23h launch) 1.93B tok
18a4cc7
Raw History Blame Contribute Delete
1.3 kB
{
"layers": 16,
"hidden": 1024,
"inter": 1536,
"heads_q": 8,
"heads_kv": 4,
"vocab": 151939,
"mask_id": 151671,
"k_heads": 8,
"n_big": 2097152,
"n_tri": 12582912,
"edim": 256,
"fusion_layers": [
1,
7
],
"teacher": "jinaai/jina-embeddings-v5-text-small-retrieval",
"phase": "C",
"objective": "cosine-1.0+mrl-each-0.3+infoNCE-learnable-1.0+gor-0.1+mlm-0.1",
"tokenizer": "jinaai/jina-embeddings-v5-text-small-retrieval",
"schedule_version": 1,
"books_epoch": 3,
"seq_len": 1024,
"seq_long": 4096,
"long_frac": 0.25,
"mix_short": {
"web": 0.5,
"books": 0.5
},
"mix_long": {
"web": 0.2,
"books": 0.8
},
"pool_mode": "dual",
"pool_queries": 8,
"pool_ablate_baseline": false,
"query_prefix": "Query:",
"doc_prefix": "Document:",
"warmup_steps": 500,
"finalize": true,
"final_decay_steps": 40000,
"planned_steps": 100000,
"engram_unfreeze_frac": 0.05,
"engram_unfreeze_step": 3863,
"base_lr": 0.00015,
"distill_lambda": 1.0,
"mrl_distill_lambda": 0.3,
"nce_lambda": 1.0,
"gor_lambda": 0.1,
"mlm_lambda": 0.1,
"log_temp_init": -2.995732273553991,
"false_neg_sim_cap": 0.95,
"bank_size": 65536,
"ance_refresh_steps": 5000,
"n_hard": 3,
"adjacent_frac": 0.5,
"min_pair_tokens": 128,
"mrl_lambda": 0.01,
"phaseB_repo": "electroglyph/berty-phaseB"
}