sync: snapshot at correct top-level path

results/current_snapshot_20260426.md

@@ -0,0 +1,123 @@
+# DNAThinker — current results snapshot (2026-04-26 00:40)
+
+Hand-aggregated from `metrics.json` files; the slower
+`build_results_table.py` will fill in once the in-flight jobs finish
+and 226046 fires.
+
+## T1 enhancer_generation grid (separatedQA, complete)
+
+`runs/exp_t1_grid_separatedQA_20260424_154915/{zs_raw,zs_enriched,lora_raw,lora_enriched}/`
+
+| Mode | parse | gc_err ↓ | len_ratio | FBD ↓ | spec ↑ | argmax ↑ | div | emb_cos |
+|---|---|---|---|---|---|---|---|---|
+| zs_raw       | 1.000 | 0.0932 | 1.829 | **11.93** | 2.752 | 0.328 | 0.442 | 0.651 |
+| zs_enriched  | 1.000 | 0.0957 | 1.615 | **11.32** | 3.129 | 0.328 | 0.430 | 0.713 |
+| lora_raw     | 1.000 | **0.0698** | 3.642 | 29.27 | 3.236 | **0.844** | 0.203 | 0.880 |
+| lora_enriched| 1.000 | 0.1023 | 3.897 | 32.50 | 2.753 | 0.578 | 0.278 | 0.837 |
+
+Notable tension: **LoRA wins gc_err + argmax_acc but loses badly on
+FBD** (29 vs 11) — it generates 2× too-long sequences and drifts from
+the real DNA distribution. Strong motivation for the Fusion-SFT →
+Loop-SFT → SV-GSPO chain.
+
+Per-cell-type FID exists in `*/genqual/genqual.json::per_cell_type` —
+currently single-cell (Ex) only since these are sample-128 evals.
+
+## T1 enhancer_generation grid (original-DEDUP, in flight as 226007)
+
+`runs/exp_t1_grid_original_DEDUP_20260425_165628/{zs_raw,zs_enriched,lora_raw,lora_enriched}/`
+
+| Mode | parse | gc_err | len_ratio | status |
+|---|---|---|---|---|
+| zs_raw | 1.000 | 0.1126 | 1.948 | ✅ |
+| zs_enriched | 1.000 | 0.1090 | 2.113 | ✅ |
+| lora_raw | 1.000 | 0.1529 | 3.717 | ✅ |
+| lora_enriched | — | — | — | 🔄 226007 in flight |
+
+Genqual not yet computed for this split.
+
+## T2 pair_aux ablation (production, n=128 each)
+
+`runs/exp_t2_pair_aux_{none,supcon_pair,tier_aware_supcon}_20260425_192434_prod/`
+
+| Variant | acc | F1 | precision | recall | parse |
+|---|---|---|---|---|---|
+| none (no aux)         | **0.773** | **0.808** | 0.701 | 0.953 | 1.000 |
+| supcon_pair           | 0.719 | 0.710 | 0.733 | 0.688 | 1.000 |
+| tier_aware_supcon     | 0.711 | 0.776 | 0.634 | 1.000 | 1.000 |
+
+⚠️ small (n=128) eval; the no-aux baseline edges out tier_aware_supcon
+on F1. The full-data run (#42, blocked on dataset rewrite 225994) will
+give the real number.
+
+## Aligner loss ablation (7-cell production)
+
+`runs/exp_aligner_t1_{infonce,lit,siglip}_20260425_210442_7cell/`
+
+| Variant | val/train ratio | Wandb |
+|---|---|---|
+| lit          | 1.22 (best generalisation gap) | dnathinker-align |
+| infonce      | 1.27 | |
+| siglip       | (in flight, 226025) | |
+
+Per memory note `reference_benchmark_suite`: **lit overfits less than
+infonce** on the 7-cell strat7c split (ratio 1.22 < 1.27).
+
+## Oracle weights
+
+| Path | Size | val_pearson_mean | val_spearman_mean |
+|---|---|---|---|
+| `runs/exp_oracle_ds_7cell_fdr_both_20260424_162210/oracle.pt` | 1.4 MB | 0.136 | 0.086 |
+| `runs/exp_oracle_ds_7cell_100k_20260424_003143/oracle.pt`     | 1.4 MB | (debug) | |
+| `runs/exp_oracle_enformer_full_<jid>/`                        | (in flight, 225956, ~30h elapsed) | | |
+
+Per-cell pearson range (DeepSTARR-7cell fdr_both): -0.017 (Mic) → 0.363 (Ast).
+
+## Currently in flight (squeue snapshot, 2026-04-26 00:40)
+
+| JID | Job | Elapsed | State |
+|---|---|---|---|
+| 225956 | oracle_enformer_full | 1d 8h | RUNNING |
+| 225994 | T2 dataset rewrite (225994) | 20h+ | RUNNING (blocks #42) |
+| 226007 | T1 task_prog (original-DEDUP) | 7h 42m | RUNNING |
+| 226025 | aligner T1 siglip 7-cell | 3h 34m | RUNNING |
+| 226037 | arch_llava (control) | 2h 14m | RUNNING (step ~110/4375) |
+| 226038 | encoder NTv3-8m | 2h 14m | RUNNING (step ~110/4375) |
+| 226043 | sv_gspo_v5 (NTv3-650m, fixed) | ~2 min | RUNNING (loading NTv3) |
+| 226044 | arch_unified_ntp (messages-fix) | ~2 min | RUNNING |
+| 226045 | arch_unified_mdlm (messages-fix) | PENDING | (Resources) |
+| 226046 | arch_ablation_table | PENDING | (Dependency afterany 226037,226038,226043,226044,226045) |
+
+## Cancelled / failed (with root cause + fix commit)
+
+| JID | Why | Fix |
+|---|---|---|
+| 226030/226031/226032 | unk_token=None Qwen3 BPE | `8acf261` (committed pre-resubmit) |
+| 226033/226034 | LoRA leaf-view in-place op | `b43c106` |
+| 226034 (resub) | SDPA mask dtype mismatch | `1c5e270` |
+| 226039 | sv_gspo encoder size mismatch (650m ckpt vs 8m model) | resubmit 226043 with NTv3-650m |
+| 226040/226041 | unified loss=0 (UnifiedCollator read flat fields, dataset emits messages) | `9f6706e` (this turn) |
+
+## Aggregator runs
+
+```bash
+# After 226037-226045 all finish (226046 auto-fires)
+results/arch_ablation_20260425_220006.md
+results/encoder_ablation_20260425.md
+
+# Manual rerun (faster after run dirs land):
+pixi run python scripts/build_results_table.py \
+    --runs-root /extra/.../runs \
+    --name-filter 'exp_t1_arch_*_20260425_220006' \
+    --output results/arch_ablation_20260425_220006.md \
+    --per-cell-type --per-tier
+```
+
+## Open questions
+
+1. Does the unified-mode messages-schema fix actually produce non-zero
+   loss now? Watching 226044 — first step's loss in ~10 min.
+2. Should we push oracle.pt + the v5 warm-start (7.2 GB) + the T1 grid
+   metrics to HF for two-machine sync? Need `HF_TOKEN`.
+3. T2 #42 still blocked on 225994's train.pair_prediction.jsonl
+   completing rewrite (~17% currently).