aether-v7.1-unified / adapter_config.json
BlockArtica's picture
Aether v7.1-unified: 10-Sephirot adapter on Qwen2.5-7B (+1.66% CE on Aether holdout, 9/9 domains, 0 regressions)
b079e9d verified
Raw
History Blame Contribute Delete
1.22 kB
{
"model_type": "sephirot_adapter",
"adapter_version": "v7.1-unified",
"framework": "candle",
"base_model": "Qwen/Qwen2.5-7B-Instruct",
"base_serve_format": "Q4_K_M GGUF (in-process, candle quantized)",
"hidden_size": 3584,
"rank": 16,
"num_experts": 10,
"expert_names": [
"Keter", "Chochmah", "Binah", "Chesed", "Gevurah",
"Tiferet", "Netzach", "Hod", "Yesod", "Malkuth"
],
"top_k": 2,
"alpha": 16.0,
"scale": 1.0,
"expert_form": "up(gelu(down(x))), up zero-initialised (identity at init)",
"trainable_params": 1182720,
"weights_dtype": "bf16",
"training": {
"objective": "plain cross-entropy (domain specialisation; base frozen)",
"corpus": "aether-curated-v3 (content-addressed Aether knowledge fabric export)",
"holdout": "aether-v6-eval-holdout (500 samples, Sephirot-domain tagged)",
"steps": 3000,
"context": 192,
"lr": 5e-4,
"optimizer": "AdamW",
"precision": "bf16",
"regression_guard": "held-out CE ceiling = base + 0.15 (PASS, base intact)"
},
"eval_full_holdout": {
"base_ce": 2.2450,
"adapter_ce": 2.2078,
"delta": -0.0373,
"relative_improvement": "1.66%",
"domains_helped": 9,
"domains_hurt": 0
}
}