anima-v3-e2 — ConsciousDecoderV3 corpus-axis fire (E2: anima 50%)
P21H V3 corpus-axis ablation, E2 variant — wiki_frac = 0.5 (anima corpus 50%, multilingual wiki 50%).
Part of the V3 fire saga (HEXAD/V3).
Verdict: FAIL
| Metric | Value |
|---|---|
| Aggregate | FAIL — 0 STRONG / 0 PARTIAL / 5 WEAK |
| per-lang | en WEAK 4/20 · ko WEAK 5/20 · zh WEAK 1/20 · ru WEAK 3/20 · ja WEAK 1/20 |
anima_register_hits |
9/20 (closure A-fire was 11/20, E3 anima-0% was 0/20) |
register_regress |
False |
| init_CE → best_CE → final_CE | 14.18 → 2.15 → 2.15 |
| train_wall | 5238 s |
Key finding
At a balanced 50/50 anima/wiki mix the anima-register memorization collapse is almost fully
back: anima_register_hits 9/20, near the 70%-anima A-fire's 11/20. The EN/KO generations
emit anima register fragments (Tension flows into this vacuum.</carve>,
진공점 [0.50,0.50], top emotion depth); ZH/RU/JA degrade into repetition loops. Halving the
anima share from 70% to 50% does not linearly relieve register collapse — V3 still FAILs.
Together with E3 (anima 0%, register hits 0/20), the corpus axis is exhausted: no anima ratio
reopens V3.
Config
- base_model:
Qwen/Qwen2.5-1.5B· init_variantqwen - n_total_params: 2,999,735,296 · vocab 151,936
- steps 5000 target — osc-detect early-stop @ step 1125 (CE mode collapse) · lr 5e-5 · bsz 2 · block_size 512 · warmup 100
- noise_sigma 0.1 · lambda_mitosis 0.0 (R2) · mitosis_max 16 (R6) · osc-detect v2.2 (thr 0.5 / window 10 / es_patience 8)
- mix: wiki_frac 0.5 (actual 0.49999) · wiki 22,073 records / 37.75 MB · anima 28,407 records / 37.75 MB · total 75.50 MB
- corpus sha256
db8f15d412817b532480522e6fb65a451856e23ab0fb9009014f4c58570e104d - mitosis: cells 2 → 16, 14 splits, 0 merges · Φ 0.712 → 0.658
Artifacts (all present, complete)
ckpt_best.pt— best checkpoint, step 1125, 6,014,446,934 bytesresult.json— full fire result (verdicts, per-lang, eval rows, mitosis summary)heldout_vp21h_v3.json— held-out evalvp21h_v3_eval1.json— anima eval1 (greedy + sample)kosmos_anchors/— 7 .kosmos emission anchors (step 500/1000/5000)train.log·dispatch.log— training + dispatch logsmix_info.json·multi_wiki_source.json·pod_id.txt— provenance
Mirror of HEXAD/UNCLASSIFIED/state/grid_3b_s187_2026_05_21/vP21H_e2/.
Visibility public — dancinlab free-tier private storage cap reached; @D a_hf_complete
(COMPLETE upload incl. 6 GB ckpt) takes precedence over the private preference for a
FAIL-verdict negative-result research checkpoint.
Model tree for dancinlab/anima-v3-e2
Base model
Qwen/Qwen2.5-1.5B