Add the one- and two-query towers (campaign 3) and the five-fold eight-query towers
Browse filesSix campaign-3 checkpoints (slot1i2cemean_icetf and slot2i2cemean_icetf, folds 0 to 2, _d512) listed in c3_docfix512_ckpt_plan.json; the five five-fold slot8i2cemean_icetf checkpoints at _g4096, at the rank rule's picks, listed in q8_g4096_ckpt_plan.json; README sections for both and the SetPoolN(Q) note.
- README.md +24 -2
- c3_docfix512_ckpt_plan.json +68 -0
- ckpt_w9cv_wcle_slot1i2cemean_icetf_fold0_d512_fp_ep1850.pt +3 -0
- ckpt_w9cv_wcle_slot1i2cemean_icetf_fold1_d512_fp_ep1850.pt +3 -0
- ckpt_w9cv_wcle_slot1i2cemean_icetf_fold2_d512_fp_ep1750.pt +3 -0
- ckpt_w9cv_wcle_slot2i2cemean_icetf_fold0_d512_fp_ep2000.pt +3 -0
- ckpt_w9cv_wcle_slot2i2cemean_icetf_fold1_d512_fp_ep900.pt +3 -0
- ckpt_w9cv_wcle_slot2i2cemean_icetf_fold2_d512_fp_ep2000.pt +3 -0
- ckpt_w9cv_wcle_slot8i2cemean_icetf_fold0_g4096_fp_ep1600.pt +3 -0
- ckpt_w9cv_wcle_slot8i2cemean_icetf_fold1_g4096_fp_ep1750.pt +3 -0
- ckpt_w9cv_wcle_slot8i2cemean_icetf_fold2_g4096_fp_ep1700.pt +3 -0
- ckpt_w9cv_wcle_slot8i2cemean_icetf_fold3_g4096_fp_ep950.pt +3 -0
- ckpt_w9cv_wcle_slot8i2cemean_icetf_fold4_g4096_fp_ep1400.pt +3 -0
- q8_g4096_ckpt_plan.json +62 -0
README.md
CHANGED
|
@@ -20,8 +20,9 @@ The data folder the towers read is the dataset
|
|
| 20 |
A file is named `ckpt_w9cv_wcle_<arm>_fold<k>[_g<cap>][_<suffix>]_fp_ep<N>.pt`: the training
|
| 21 |
arm, the fold of the five-fold split, the anchor budget it trained with (none = 512), and the
|
| 22 |
epoch. Each file is a PyTorch state dict of `Pod/w9_cv_worker.py`'s `SetPoolN(4)` (keys `q0`,
|
| 23 |
-
`attn.*`, `head.*`)
|
| 24 |
-
|
|
|
|
| 25 |
|
| 26 |
Score them against the manuscript's ten 4,096-sentence anchor draws (from the repository root):
|
| 27 |
|
|
@@ -95,3 +96,24 @@ packs were redone. Selected Stripped hit@1 on the ten 4,096-sentence draws (five
|
|
| 95 |
| τ = 0.05 | 0.601 ± 0.029 | 0.642 ± 0.026 |
|
| 96 |
| τ = 0.10 | 0.510 ± 0.022 (four folds) | 0.520 ± 0.015 |
|
| 97 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
A file is named `ckpt_w9cv_wcle_<arm>_fold<k>[_g<cap>][_<suffix>]_fp_ep<N>.pt`: the training
|
| 21 |
arm, the fold of the five-fold split, the anchor budget it trained with (none = 512), and the
|
| 22 |
epoch. Each file is a PyTorch state dict of `Pod/w9_cv_worker.py`'s `SetPoolN(4)` (keys `q0`,
|
| 23 |
+
`attn.*`, `head.*`), or of `SetPoolN(Q)` for the `slot<Q>i2cemean_icetf` arms, whose `q0` has shape
|
| 24 |
+
[1, Q, 128]; the VICReg runs hold `{"model": ..., "expander": ...}`. Only tensors, no code objects,
|
| 25 |
+
so `torch.load(path, weights_only=True)` reads them.
|
| 26 |
|
| 27 |
Score them against the manuscript's ten 4,096-sentence anchor draws (from the repository root):
|
| 28 |
|
|
|
|
| 96 |
| τ = 0.05 | 0.601 ± 0.029 | 0.642 ± 0.026 |
|
| 97 |
| τ = 0.10 | 0.510 ± 0.022 (four folds) | 0.520 ± 0.015 |
|
| 98 |
|
| 99 |
+
## One and two latent queries (campaign 3), added 2026-11-10
|
| 100 |
+
|
| 101 |
+
The latent-query sweep's one- and two-query towers (`slot1i2cemean_icetf`, `slot2i2cemean_icetf`),
|
| 102 |
+
folds 0 to 2, trained on the corrected packs under campaign 2's recipe and suffix: six runs, the
|
| 103 |
+
selected checkpoint of each listed in `c3_docfix512_ckpt_plan.json` (campaign 2's entry format).
|
| 104 |
+
With the four-query I-CE of `docfix512_ckpt_plan.json` and campaign 2's 8 and 16 queries they give
|
| 105 |
+
the sweep's five query counts at 512.
|
| 106 |
+
|
| 107 |
+
## Eight latent queries, five folds (`_g4096`), added 2026-11-10
|
| 108 |
+
|
| 109 |
+
The five-fold eight-query tower (`slot8i2cemean_icetf`), trained at the anchor cap of the four-query
|
| 110 |
+
I-CE runs it is compared with (`_g4096`). Its packs were built from the 2,048-sentence review pool
|
| 111 |
+
(the training logs show `anchors: used med 2068`), so the 512-sentence correction does not touch it.
|
| 112 |
+
The worker's own pick (`best_ep` in its `zsbest` record) came from a different selection rule; the
|
| 113 |
+
five checkpoints in `q8_g4096_ckpt_plan.json` are the manuscript's rank rule applied to the towers
|
| 114 |
+
the worker saved with every checkpoint, the procedure that reproduces `rank_ckpt_plan.json`'s picks
|
| 115 |
+
for the four-query runs. Campaign 2's entry format, plus `cap`.
|
| 116 |
+
|
| 117 |
+
`k10_eval.py`'s `load_tower` builds `SetPoolN(4)`; to score a `slot<Q>` file, build `SetPoolN(Q)`
|
| 118 |
+
from the shape of its `q0` instead.
|
| 119 |
+
|
c3_docfix512_ckpt_plan.json
ADDED
|
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"slot1i2cemean_icetf_fold0_d512": {
|
| 3 |
+
"arm": "slot1i2cemean_icetf",
|
| 4 |
+
"fold": 0,
|
| 5 |
+
"protocol": "five-fold (seed = fold)",
|
| 6 |
+
"worker": "Pod/w9_cv_worker.py",
|
| 7 |
+
"sfx": "_d512",
|
| 8 |
+
"ep": 1850,
|
| 9 |
+
"ckpt": "ckpt_w9cv_wcle_slot1i2cemean_icetf_fold0_d512_fp_ep1850.pt",
|
| 10 |
+
"sha256": "7a0c9fa191f2277833a076d6260c998c5cc439322ed956785e69db2561fca22e",
|
| 11 |
+
"ranksel": 102.129047782017
|
| 12 |
+
},
|
| 13 |
+
"slot1i2cemean_icetf_fold1_d512": {
|
| 14 |
+
"arm": "slot1i2cemean_icetf",
|
| 15 |
+
"fold": 1,
|
| 16 |
+
"protocol": "five-fold (seed = fold)",
|
| 17 |
+
"worker": "Pod/w9_cv_worker.py",
|
| 18 |
+
"sfx": "_d512",
|
| 19 |
+
"ep": 1850,
|
| 20 |
+
"ckpt": "ckpt_w9cv_wcle_slot1i2cemean_icetf_fold1_d512_fp_ep1850.pt",
|
| 21 |
+
"sha256": "cc6a92592ac81d681637c390e94dba57f5a1e0fb28b4cd295a6b743bf8f2c16e",
|
| 22 |
+
"ranksel": 98.1664812043291
|
| 23 |
+
},
|
| 24 |
+
"slot1i2cemean_icetf_fold2_d512": {
|
| 25 |
+
"arm": "slot1i2cemean_icetf",
|
| 26 |
+
"fold": 2,
|
| 27 |
+
"protocol": "five-fold (seed = fold)",
|
| 28 |
+
"worker": "Pod/w9_cv_worker.py",
|
| 29 |
+
"sfx": "_d512",
|
| 30 |
+
"ep": 1750,
|
| 31 |
+
"ckpt": "ckpt_w9cv_wcle_slot1i2cemean_icetf_fold2_d512_fp_ep1750.pt",
|
| 32 |
+
"sha256": "37456aabdff722f190d7e914ff7bdd714180f777bfd100278b91e2f10791fd39",
|
| 33 |
+
"ranksel": 103.26830974538323
|
| 34 |
+
},
|
| 35 |
+
"slot2i2cemean_icetf_fold0_d512": {
|
| 36 |
+
"arm": "slot2i2cemean_icetf",
|
| 37 |
+
"fold": 0,
|
| 38 |
+
"protocol": "five-fold (seed = fold)",
|
| 39 |
+
"worker": "Pod/w9_cv_worker.py",
|
| 40 |
+
"sfx": "_d512",
|
| 41 |
+
"ep": 2000,
|
| 42 |
+
"ckpt": "ckpt_w9cv_wcle_slot2i2cemean_icetf_fold0_d512_fp_ep2000.pt",
|
| 43 |
+
"sha256": "2beba6ddd569fa1d31f645c9c9da2f732b94bd902c40461946a392ae6e640f43",
|
| 44 |
+
"ranksel": 101.4992104596883
|
| 45 |
+
},
|
| 46 |
+
"slot2i2cemean_icetf_fold1_d512": {
|
| 47 |
+
"arm": "slot2i2cemean_icetf",
|
| 48 |
+
"fold": 1,
|
| 49 |
+
"protocol": "five-fold (seed = fold)",
|
| 50 |
+
"worker": "Pod/w9_cv_worker.py",
|
| 51 |
+
"sfx": "_d512",
|
| 52 |
+
"ep": 900,
|
| 53 |
+
"ckpt": "ckpt_w9cv_wcle_slot2i2cemean_icetf_fold1_d512_fp_ep900.pt",
|
| 54 |
+
"sha256": "cca438d55807af9e124c9ade4ca15b54054cf1ce1226f861e33c044abf8073a9",
|
| 55 |
+
"ranksel": 99.15674234783602
|
| 56 |
+
},
|
| 57 |
+
"slot2i2cemean_icetf_fold2_d512": {
|
| 58 |
+
"arm": "slot2i2cemean_icetf",
|
| 59 |
+
"fold": 2,
|
| 60 |
+
"protocol": "five-fold (seed = fold)",
|
| 61 |
+
"worker": "Pod/w9_cv_worker.py",
|
| 62 |
+
"sfx": "_d512",
|
| 63 |
+
"ep": 2000,
|
| 64 |
+
"ckpt": "ckpt_w9cv_wcle_slot2i2cemean_icetf_fold2_d512_fp_ep2000.pt",
|
| 65 |
+
"sha256": "7fe02695f5ac123273017d39cfd70ac9bc5795f028235c2397bf5fb1d850255c",
|
| 66 |
+
"ranksel": 103.84762984590806
|
| 67 |
+
}
|
| 68 |
+
}
|
ckpt_w9cv_wcle_slot1i2cemean_icetf_fold0_d512_fp_ep1850.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7a0c9fa191f2277833a076d6260c998c5cc439322ed956785e69db2561fca22e
|
| 3 |
+
size 1451122
|
ckpt_w9cv_wcle_slot1i2cemean_icetf_fold1_d512_fp_ep1850.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cc6a92592ac81d681637c390e94dba57f5a1e0fb28b4cd295a6b743bf8f2c16e
|
| 3 |
+
size 1451122
|
ckpt_w9cv_wcle_slot1i2cemean_icetf_fold2_d512_fp_ep1750.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:37456aabdff722f190d7e914ff7bdd714180f777bfd100278b91e2f10791fd39
|
| 3 |
+
size 1451122
|
ckpt_w9cv_wcle_slot2i2cemean_icetf_fold0_d512_fp_ep2000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2beba6ddd569fa1d31f645c9c9da2f732b94bd902c40461946a392ae6e640f43
|
| 3 |
+
size 1451634
|
ckpt_w9cv_wcle_slot2i2cemean_icetf_fold1_d512_fp_ep900.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cca438d55807af9e124c9ade4ca15b54054cf1ce1226f861e33c044abf8073a9
|
| 3 |
+
size 1451617
|
ckpt_w9cv_wcle_slot2i2cemean_icetf_fold2_d512_fp_ep2000.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7fe02695f5ac123273017d39cfd70ac9bc5795f028235c2397bf5fb1d850255c
|
| 3 |
+
size 1451634
|
ckpt_w9cv_wcle_slot8i2cemean_icetf_fold0_g4096_fp_ep1600.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8c4ac1e5cec2c68cb5a3cfd7abf6d56beb29b325c3b30cc9e92056d46d1e2689
|
| 3 |
+
size 1454228
|
ckpt_w9cv_wcle_slot8i2cemean_icetf_fold1_g4096_fp_ep1750.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:912bb1c0fb996b1d388e661777666a75f350e9a2c7b4442e3497030e21598a33
|
| 3 |
+
size 1454228
|
ckpt_w9cv_wcle_slot8i2cemean_icetf_fold2_g4096_fp_ep1700.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c025e568fae245ca95d837f299086f76ad86481c746e9afffced4f1d75cf14fd
|
| 3 |
+
size 1454228
|
ckpt_w9cv_wcle_slot8i2cemean_icetf_fold3_g4096_fp_ep950.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:783867378801212e7efc5c47f523e5a812acc2cc0403276b7f3e88e091155584
|
| 3 |
+
size 1454213
|
ckpt_w9cv_wcle_slot8i2cemean_icetf_fold4_g4096_fp_ep1400.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:02348eb250e4b36fcc22cdbfa6b183bbe551e001444c221a68adba5bb6e1023a
|
| 3 |
+
size 1454228
|
q8_g4096_ckpt_plan.json
ADDED
|
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"slot8i2cemean_icetf_fold0_g4096": {
|
| 3 |
+
"arm": "slot8i2cemean_icetf",
|
| 4 |
+
"fold": 0,
|
| 5 |
+
"cap": "_g4096",
|
| 6 |
+
"protocol": "five-fold (seed = fold)",
|
| 7 |
+
"worker": "Pod/w9_cv_worker.py",
|
| 8 |
+
"sfx": "",
|
| 9 |
+
"ep": 1600,
|
| 10 |
+
"ckpt": "ckpt_w9cv_wcle_slot8i2cemean_icetf_fold0_g4096_fp_ep1600.pt",
|
| 11 |
+
"sha256": "8c4ac1e5cec2c68cb5a3cfd7abf6d56beb29b325c3b30cc9e92056d46d1e2689",
|
| 12 |
+
"ranksel": 106.1260102083505
|
| 13 |
+
},
|
| 14 |
+
"slot8i2cemean_icetf_fold1_g4096": {
|
| 15 |
+
"arm": "slot8i2cemean_icetf",
|
| 16 |
+
"fold": 1,
|
| 17 |
+
"cap": "_g4096",
|
| 18 |
+
"protocol": "five-fold (seed = fold)",
|
| 19 |
+
"worker": "Pod/w9_cv_worker.py",
|
| 20 |
+
"sfx": "",
|
| 21 |
+
"ep": 1750,
|
| 22 |
+
"ckpt": "ckpt_w9cv_wcle_slot8i2cemean_icetf_fold1_g4096_fp_ep1750.pt",
|
| 23 |
+
"sha256": "912bb1c0fb996b1d388e661777666a75f350e9a2c7b4442e3497030e21598a33",
|
| 24 |
+
"ranksel": 106.16756746064351
|
| 25 |
+
},
|
| 26 |
+
"slot8i2cemean_icetf_fold2_g4096": {
|
| 27 |
+
"arm": "slot8i2cemean_icetf",
|
| 28 |
+
"fold": 2,
|
| 29 |
+
"cap": "_g4096",
|
| 30 |
+
"protocol": "five-fold (seed = fold)",
|
| 31 |
+
"worker": "Pod/w9_cv_worker.py",
|
| 32 |
+
"sfx": "",
|
| 33 |
+
"ep": 1700,
|
| 34 |
+
"ckpt": "ckpt_w9cv_wcle_slot8i2cemean_icetf_fold2_g4096_fp_ep1700.pt",
|
| 35 |
+
"sha256": "c025e568fae245ca95d837f299086f76ad86481c746e9afffced4f1d75cf14fd",
|
| 36 |
+
"ranksel": 106.87001435083803
|
| 37 |
+
},
|
| 38 |
+
"slot8i2cemean_icetf_fold3_g4096": {
|
| 39 |
+
"arm": "slot8i2cemean_icetf",
|
| 40 |
+
"fold": 3,
|
| 41 |
+
"cap": "_g4096",
|
| 42 |
+
"protocol": "five-fold (seed = fold)",
|
| 43 |
+
"worker": "Pod/w9_cv_worker.py",
|
| 44 |
+
"sfx": "",
|
| 45 |
+
"ep": 950,
|
| 46 |
+
"ckpt": "ckpt_w9cv_wcle_slot8i2cemean_icetf_fold3_g4096_fp_ep950.pt",
|
| 47 |
+
"sha256": "783867378801212e7efc5c47f523e5a812acc2cc0403276b7f3e88e091155584",
|
| 48 |
+
"ranksel": 106.50847710665332
|
| 49 |
+
},
|
| 50 |
+
"slot8i2cemean_icetf_fold4_g4096": {
|
| 51 |
+
"arm": "slot8i2cemean_icetf",
|
| 52 |
+
"fold": 4,
|
| 53 |
+
"cap": "_g4096",
|
| 54 |
+
"protocol": "five-fold (seed = fold)",
|
| 55 |
+
"worker": "Pod/w9_cv_worker.py",
|
| 56 |
+
"sfx": "",
|
| 57 |
+
"ep": 1400,
|
| 58 |
+
"ckpt": "ckpt_w9cv_wcle_slot8i2cemean_icetf_fold4_g4096_fp_ep1400.pt",
|
| 59 |
+
"sha256": "02348eb250e4b36fcc22cdbfa6b183bbe551e001444c221a68adba5bb6e1023a",
|
| 60 |
+
"ranksel": 104.94705536826554
|
| 61 |
+
}
|
| 62 |
+
}
|