chironYu commited on
Commit
3f0f91a
·
verified ·
1 Parent(s): 3f26346

Card and selection files for the checkpoints added today; the citation names the first author

Browse files

rank_ckpt_plan.json and fixed_split_ckpt_plan.json are the files of contrast_experiment/w9/eval/ of the code repository (commit 31ce099). They add the DINO-style control and eight fixed-split runs, and list the 52 further checkpoints of two runs under `trajectory`; no earlier entry changed.

Files changed (3) hide show
  1. README.md +61 -20
  2. fixed_split_ckpt_plan.json +391 -0
  3. rank_ckpt_plan.json +11 -0
README.md CHANGED
@@ -24,13 +24,13 @@ reviews (Qwen3-Embedding-0.6B, 1,024-d) and an MLP maps the pooled result to one
24
 
25
  ## Contents
26
 
27
- 204 checkpoints, about 320 MiB, and four JSON files that list every one of them:
28
 
29
  | Listed in | Checkpoints | What they are |
30
  |---|---|---|
31
- | `rank_ckpt_plan.json` | 156 | the five-fold runs the paper's numbers read (all but one control, see below), each at the epoch the validation rank rule of Section 5 selects |
32
  | `i2bce_sel_pool.json` | 5 | I + SimCLR on the five folds (Figure 8), selected by the same rule |
33
- | `fixed_split_ckpt_plan.json` | 8 | the fixed-split runs of Table 3 and Section 6.2 that train with 512-sentence anchors: seven runs, one of them with an earlier checkpoint kept beside the selected one |
34
  | `rank_ckpt_plan.json`, field `init_ckpt` | 5 | the BYOL checkpoints the five two-stage runs started from |
35
  | `superseded_ckpt_plan.json` | 30 | kept for the record; no number of the paper reads them |
36
 
@@ -39,10 +39,11 @@ the file name and the file's SHA-256; `i2bce_sel_pool.json` is the selection as
39
  script wrote it, with the rank score of every checkpoint of the run. The same four files are in
40
  `contrast_experiment/w9/eval/` of the code repository, where the evaluation script reads them.
41
 
42
- Not in this repository: the other fixed-split towers the paper reads (the a+u @ views row of
43
- Table 3, the VICReg rows of Table A7, and the towers of Figure 9, Figures A1 and A3 and the
44
- neighborhood statistics of Appendix L), and the DINO-style control of Section 4.3 (Table A1, one
45
- fold), whose training code is not part of the code release either.
 
46
 
47
  ## File names
48
 
@@ -82,29 +83,50 @@ row reads at least one of the row's runs. That file also names the notebook that
82
  | `ce_cetf_gpool`, `i2ce_icetf_gpool` | CE; I-CE, the recipe | pool | 5 | Figure 7, Figure 8, Table 2, Section 6.1, Table A1, Table A2, Table A13, Table A14, Table A15, Table A16, Table A17, Figure A2, Table S2, Table S3, Table S4 |
83
  | `slot8i2cemean_icetf_gpool` | I-CE, 8 latent queries | pool | 5 | Table A6 |
84
  | `swin168step84loop2i2ce_icetf_gpool`, `swin168step84loop2i2ce_icetf_gpool_bw` | I-CE with the windowed teacher; I-CE, two-stage: BYOL warm start, then the windowed teacher | pool | 5 | Table A12, Table S2, Table S3, Table S4 |
 
85
 
86
  I + SimCLR (`i2bce_icetf`, files without a pack or a suffix, five folds) backs Figure 8. The
87
- fold-0 run of `i2ce_icetf_g2048` also gives the game vectors of Figure 10.
 
 
88
 
89
  ## The fixed-split runs
90
 
91
  | Runs | What they are | Pack | Folds | Read by |
92
  |---|---|---|---|---|
93
- | `ai25auni2_icetf_d512`, `ai2auni25_icetf_d512`, `ai4auni2_icetf_d512`, `ai6auni2_icetf_d512` | alignment + uniformity aimed at the anchor gallery, the four settings of Section 6.2 (`ai25auni2` is Table 3's a+u @ gallery row) | 512 | fixed split | Table 3, Section 6.2 |
94
  | `ai2ce_icetf_d512`, `ce_cetf_d512`, `i2ce_icetf_d512` | CE + anchor-in-I; CE; I-CE, the recipe | 512 | fixed split | Table 3, Section 6.2, Table A7 |
95
-
96
- The fixed-split CE + anchor-in-I run stopped at 1,000 epochs under its first schedule and was
97
- trained on to 2,000 (the weights of epoch 1,000 with a fresh optimizer); the rank rule over all
98
- 40 checkpoints selects epoch 1,650. The epoch-1,000 file is kept and the plan entry names it
99
- under `previous`.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
100
 
101
  ## Format
102
 
103
  Every file holds tensors only, so `torch.load(path, weights_only=True)` reads it. Most are the
104
  state dict of the tower itself, `SetPoolN(Q)` of `contrast_experiment/w9/Pod/w9_cv_worker.py`
105
  (keys `q0`, `attn.*`, `head.*`; `q0` has shape `[1, Q, 128]`, Q = 4 except in the `slot<Q>` runs).
106
- Three families wrap it: BYOL files hold `{"model", "pred", "target"}`, VICReg files
107
- `{"model", "expander"}`, and the memory-bank files `{"model", "shadow"}`; the tower is `model`.
 
 
 
108
 
109
  ```python
110
  import sys, torch
@@ -135,11 +157,12 @@ python contrast_experiment/w9/eval/k10_eval.py score cells.json draws.npz score
135
  python contrast_experiment/w9/eval/k10_eval.py table scores.json
136
  ```
137
 
138
- `cells` takes the 161 five-fold checkpoints of `rank_ckpt_plan.json` and `i2bce_sel_pool.json`;
139
  `--only <regex>` keeps a subset, for example `--only "_d512$"` for the 512-sentence runs. The
140
  script does not score the fixed-split towers. In its table the five families of the main
141
  comparison go by the paper's names (`I-CE`, `CE`, `SimCLR`, `VICReg`, `BYOL`) and every other
142
- family by its plan key without the fold.
 
143
 
144
  ## The 512-sentence packs and the `_d512` runs
145
 
@@ -154,6 +177,7 @@ repository give a run on the current packs that name themselves.
154
  Runs whose loss never reads the anchors (the SimCLR-style tower, I + SimCLR, BYOL, VICReg and
155
  alignment plus uniformity aimed at the batch views) train the same weights on either pack, so
156
  their checkpoints were not replaced; only readings that used the 512-sentence packs were redone.
 
157
 
158
  Stripped hit@1 on the ten 4,096-sentence draws, five-fold mean ± std, before and after:
159
 
@@ -184,6 +208,11 @@ dataset keeps the first packs as `wscan_gal_rev_rawprefix.npz`; the workers read
184
 
185
  ## Changes
186
 
 
 
 
 
 
187
  - 2026-11-11. The four JSON files above replace the earlier per-campaign lists
188
  (`docfix512_ckpt_plan.json`, `c2_docfix512_ckpt_plan.json`, `c3_docfix512_ckpt_plan.json`,
189
  `q8_gpool_ckpt_plan.json`); no selection changed. The five BYOL checkpoints the two-stage runs
@@ -193,5 +222,17 @@ dataset keeps the first packs as `wscan_gal_rev_rawprefix.npz`; the workers read
193
 
194
  ## Citation
195
 
196
- *Invariant Game Representation Learning from Steam Reviews.* AICCC 2026. Code and checkpoints are
197
- under the MIT license.
 
 
 
 
 
 
 
 
 
 
 
 
 
24
 
25
  ## Contents
26
 
27
+ 265 checkpoints, about 580 MiB, and four JSON files that list every one of them:
28
 
29
  | Listed in | Checkpoints | What they are |
30
  |---|---|---|
31
+ | `rank_ckpt_plan.json` | 157 | the five-fold runs the paper's numbers read, each at the epoch the validation rank rule of Section 5 selects; one of them is the DINO-style control of Table A1, trained on fold 0 only |
32
  | `i2bce_sel_pool.json` | 5 | I + SimCLR on the five folds (Figure 8), selected by the same rule |
33
+ | `fixed_split_ckpt_plan.json` | 68 | the fixed-split runs: 15 runs at their selected epochs, one of them with an earlier checkpoint kept beside the selected one, and 52 more checkpoints of two runs that the paper reads across their epochs (field `trajectory`) |
34
  | `rank_ckpt_plan.json`, field `init_ckpt` | 5 | the BYOL checkpoints the five two-stage runs started from |
35
  | `superseded_ckpt_plan.json` | 30 | kept for the record; no number of the paper reads them |
36
 
 
39
  script wrote it, with the rank score of every checkpoint of the run. The same four files are in
40
  `contrast_experiment/w9/eval/` of the code repository, where the evaluation script reads them.
41
 
42
+ Every tower that a number or a figure of the paper reads is here, at the checkpoint it is read
43
+ at. A run saved a checkpoint every 50 epochs and the rank rule chose among them; the ones it did
44
+ not choose are not here, except for the two runs listed under `trajectory`. The one other passage
45
+ that follows runs over training, the view-count sweep at early epochs in Appendix C, quotes the
46
+ readings the workers logged at each checkpoint.
47
 
48
  ## File names
49
 
 
83
  | `ce_cetf_gpool`, `i2ce_icetf_gpool` | CE; I-CE, the recipe | pool | 5 | Figure 7, Figure 8, Table 2, Section 6.1, Table A1, Table A2, Table A13, Table A14, Table A15, Table A16, Table A17, Figure A2, Table S2, Table S3, Table S4 |
84
  | `slot8i2cemean_icetf_gpool` | I-CE, 8 latent queries | pool | 5 | Table A6 |
85
  | `swin168step84loop2i2ce_icetf_gpool`, `swin168step84loop2i2ce_icetf_gpool_bw` | I-CE with the windowed teacher; I-CE, two-stage: BYOL warm start, then the windowed teacher | pool | 5 | Table A12, Table S2, Table S3, Table S4 |
86
+ | `dino_dntf_gpool` | DINO at its published settings, the control of Section 4.3 | pool | 0 only | Table A1 |
87
 
88
  I + SimCLR (`i2bce_icetf`, files without a pack or a suffix, five folds) backs Figure 8. The
89
+ fold-0 run of `i2ce_icetf_g2048` also gives the game vectors of Figure 10. The DINO-style control
90
+ has one fold, so Table A1 gives its spread over the ten anchor draws; its file is 169 MiB because
91
+ it carries the two prototype heads (see Format).
92
 
93
  ## The fixed-split runs
94
 
95
  | Runs | What they are | Pack | Folds | Read by |
96
  |---|---|---|---|---|
97
+ | `ai25auni2_icetf_d512`, `ai2auni25_icetf_d512`, `ai4auni2_icetf_d512`, `ai6auni2_icetf_d512`, `i4uni2_icetf` | alignment + uniformity aimed at the anchor gallery, the four settings of Section 6.2 (`ai25auni2` is Table 3's a+u @ gallery row); alignment + uniformity aimed at the batch views (Table 3's a+u @ views row) | 512 | fixed split | Table 3, Section 6.2 |
98
  | `ai2ce_icetf_d512`, `ce_cetf_d512`, `i2ce_icetf_d512` | CE + anchor-in-I; CE; I-CE, the recipe | 512 | fixed split | Table 3, Section 6.2, Table A7 |
99
+ | `ce_cetf_g2048`, `i2ce_icetf_g2048` | CE; I-CE, the recipe | 2,048 | fixed split | Figure 9, Figure A1, Figure A3, Appendix L |
100
+ | `bce_cetf`, `i2bce_icetf`, `epd_v20i10c20_cetf` | SimCLR-style; I + SimCLR; VICReg, v/i/c = 20/10/20 (the recipe weights) | 512 | fixed split | Figure 9, Figure A3, Appendix L, Table A7 |
101
+ | `epd_v25i25c1_cetf`, `epd_v20i10c15_cetf` | VICReg, v/i/c = 25/25/1; VICReg, v/i/c = 20/10/15 | 512 | fixed split | Table A7 |
102
+
103
+ Figure 9, Figure A3 and the neighborhood statistics of Appendix L read six encoders on the fixed
104
+ split: the frozen embedder, which has no checkpoint, and five towers, namely `i2ce_icetf_g2048`
105
+ (epoch 1,750), `ce_cetf_g2048` (1,150), `bce_cetf` (1,350), `i2bce_icetf` (1,350) and
106
+ `epd_v20i10c20_cetf` (100). Two runs have more than their selected checkpoint here, listed under
107
+ `trajectory` in the plan file: all 40 checkpoints of `i2ce_icetf_g2048`, whose four latent queries
108
+ Figure A1 draws at every one, and the `bce_cetf` checkpoints from epoch 1,350 to 2,000, the span
109
+ of the test curve Appendix L describes.
110
+
111
+ Runs that read no anchor in training (`bce_cetf`, `i2bce_icetf`, `i4uni2_icetf` and the three
112
+ VICReg weightings; the 512 in their rows is only the pack their worker encoded at each
113
+ checkpoint) were selected after training: each of the run's 40 checkpoints encodes the
114
+ 4,096-sentence anchor pack of the fixed split and the rank rule chooses on that. These six runs
115
+ and the fixed-split CE + anchor-in-I run stopped at 1,000 epochs under their first schedule and
116
+ were trained on to 2,000 from the weights of epoch 1,000 with a fresh optimizer; the plan entries
117
+ say so under `extended`. For CE + anchor-in-I the rule then selects epoch 1,650, and the
118
+ epoch-1,000 file is kept and named under `previous`.
119
 
120
  ## Format
121
 
122
  Every file holds tensors only, so `torch.load(path, weights_only=True)` reads it. Most are the
123
  state dict of the tower itself, `SetPoolN(Q)` of `contrast_experiment/w9/Pod/w9_cv_worker.py`
124
  (keys `q0`, `attn.*`, `head.*`; `q0` has shape `[1, Q, 128]`, Q = 4 except in the `slot<Q>` runs).
125
+ Four families wrap it: BYOL files hold `{"model", "pred", "target"}`, VICReg files
126
+ `{"model", "expander"}`, the memory-bank files `{"model", "shadow"}`, and the DINO-style file
127
+ `{"model", "student", "t_head", "s_head", "center"}`; the tower is `model`. In the DINO-style file
128
+ `model` is the teacher's tower, the one DINO evaluates and the paper reads, `student` the
129
+ student's, and the two heads hold the 65,536 prototypes.
130
 
131
  ```python
132
  import sys, torch
 
157
  python contrast_experiment/w9/eval/k10_eval.py table scores.json
158
  ```
159
 
160
+ `cells` takes the 162 five-fold checkpoints of `rank_ckpt_plan.json` and `i2bce_sel_pool.json`;
161
  `--only <regex>` keeps a subset, for example `--only "_d512$"` for the 512-sentence runs. The
162
  script does not score the fixed-split towers. In its table the five families of the main
163
  comparison go by the paper's names (`I-CE`, `CE`, `SimCLR`, `VICReg`, `BYOL`) and every other
164
+ family by its plan key without the fold; a family with one fold, the DINO-style control, is
165
+ printed with its spread over the ten draws.
166
 
167
  ## The 512-sentence packs and the `_d512` runs
168
 
 
177
  Runs whose loss never reads the anchors (the SimCLR-style tower, I + SimCLR, BYOL, VICReg and
178
  alignment plus uniformity aimed at the batch views) train the same weights on either pack, so
179
  their checkpoints were not replaced; only readings that used the 512-sentence packs were redone.
180
+ The fixed-split files of those runs therefore carry no `_d512`.
181
 
182
  Stripped hit@1 on the ten 4,096-sentence draws, five-fold mean ± std, before and after:
183
 
 
208
 
209
  ## Changes
210
 
211
+ - 2026-11-11, later that day. 61 checkpoints were added, so that every tower the paper reads is
212
+ here: the DINO-style control (fold 0), eight fixed-split runs at their selected epochs (CE and
213
+ I-CE trained at 2,048 sentences, the SimCLR-style tower, I + SimCLR, a+u @ views and the three
214
+ VICReg weightings), and the 52 `trajectory` checkpoints. `rank_ckpt_plan.json` and
215
+ `fixed_split_ckpt_plan.json` name them; no earlier entry changed.
216
  - 2026-11-11. The four JSON files above replace the earlier per-campaign lists
217
  (`docfix512_ckpt_plan.json`, `c2_docfix512_ckpt_plan.json`, `c3_docfix512_ckpt_plan.json`,
218
  `q8_gpool_ckpt_plan.json`); no selection changed. The five BYOL checkpoints the two-stage runs
 
222
 
223
  ## Citation
224
 
225
+ Zhilong Yu et al. *Invariant Game Representation Learning from Steam Reviews.* 2026 9th Artificial
226
+ Intelligence and Cloud Computing Conference (AICCC 2026), December 18 to 21, 2026, Tokyo, Japan.
227
+
228
+ ```bibtex
229
+ @inproceedings{yu2026invariant,
230
+ author = {Yu, Zhilong and others},
231
+ title = {Invariant Game Representation Learning from Steam Reviews},
232
+ booktitle = {2026 9th Artificial Intelligence and Cloud Computing Conference (AICCC 2026)},
233
+ address = {Tokyo, Japan},
234
+ year = {2026}
235
+ }
236
+ ```
237
+
238
+ Code and checkpoints are under the MIT license.
fixed_split_ckpt_plan.json CHANGED
@@ -89,5 +89,396 @@
89
  "ranksel": 130.45839742717456,
90
  "protocol": "fixed split (seed 0)",
91
  "worker": "Pod/w9_a100_worker.py"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
92
  }
93
  }
 
89
  "ranksel": 130.45839742717456,
90
  "protocol": "fixed split (seed 0)",
91
  "worker": "Pod/w9_a100_worker.py"
92
+ },
93
+ "ce_cetf_fixed_g2048": {
94
+ "arm": "ce_cetf",
95
+ "fold": null,
96
+ "cap": "_g2048",
97
+ "sfx": "",
98
+ "ep": 1150,
99
+ "ckpt": "ckpt_w9_wcle_ce_cetf_g2048_fp_ep1150.pt",
100
+ "sha256": "63af5f9511807e489524dde74556bb6d48dd03d5ef04f66647deea7e3f43279c",
101
+ "ranksel": 125.32073730529692,
102
+ "protocol": "fixed split (seed 0)",
103
+ "worker": "Pod/w9_a100_worker.py",
104
+ "selected_on": "rank rule on the gallery the worker saved with each of the 40 checkpoints",
105
+ "reads": "Figure 9, Figure A3, Appendix L"
106
+ },
107
+ "i2ce_icetf_fixed_g2048": {
108
+ "arm": "i2ce_icetf",
109
+ "fold": null,
110
+ "cap": "_g2048",
111
+ "sfx": "",
112
+ "ep": 1750,
113
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1750.pt",
114
+ "sha256": "7459ba39465ed1b79ea6ce3380b7ad82073439dbf390e9c8b659cf41a5f5bb7a",
115
+ "ranksel": 131.9001247718065,
116
+ "protocol": "fixed split (seed 0)",
117
+ "worker": "Pod/w9_a100_worker.py",
118
+ "selected_on": "rank rule on the gallery the worker saved with each of the 40 checkpoints",
119
+ "reads": "Figure 9, Figure A1, Figure A3, Appendix L",
120
+ "trajectory": {
121
+ "why": "Figure A1 draws the four latent queries of this run at every saved checkpoint",
122
+ "files": [
123
+ {
124
+ "ep": 50,
125
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep50.pt",
126
+ "sha256": "e56eb29df5fd429d9128b888a372971a8c5f614c9f25ba675f092ac4fc92d448"
127
+ },
128
+ {
129
+ "ep": 100,
130
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep100.pt",
131
+ "sha256": "33f8c3371dff194c33ef4b2150a7c0249bc0a69c56e8dfc412759c08b4e2073a"
132
+ },
133
+ {
134
+ "ep": 150,
135
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep150.pt",
136
+ "sha256": "b4f98bce196c6c7fc78ce86934de802bff8dab7696fd5ea06650fbb07b1ab44f"
137
+ },
138
+ {
139
+ "ep": 200,
140
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep200.pt",
141
+ "sha256": "2b090e9f38c550931ef8ecfcd5c8af9a4fe907eddc4ef97eb03935d6994ef59a"
142
+ },
143
+ {
144
+ "ep": 250,
145
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep250.pt",
146
+ "sha256": "9cb42878ea46ff079d950faf9b135838ef3a531c28711ef5884fe01b11e46573"
147
+ },
148
+ {
149
+ "ep": 300,
150
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep300.pt",
151
+ "sha256": "fbe598c032383434a0a2b66bfc2e24d2af01c26d7c6e327a6b7af6569644a110"
152
+ },
153
+ {
154
+ "ep": 350,
155
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep350.pt",
156
+ "sha256": "a44e39ae4261a080f0db07df749d771cf2164a40deeb6e30238f744dbc32eefb"
157
+ },
158
+ {
159
+ "ep": 400,
160
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep400.pt",
161
+ "sha256": "7c91b7f01d973664086a7d22efb682ebf1808e7bdd0c19a548f4f03547a3c7c5"
162
+ },
163
+ {
164
+ "ep": 450,
165
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep450.pt",
166
+ "sha256": "fe07f155470b4a43c08571cc42a9b390771908371877a4c00baab195a317edc2"
167
+ },
168
+ {
169
+ "ep": 500,
170
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep500.pt",
171
+ "sha256": "5317180a051c479f5b4e11f3dc7a821569e9f514c2b4189934c5b7a8505cdd82"
172
+ },
173
+ {
174
+ "ep": 550,
175
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep550.pt",
176
+ "sha256": "66dc84cf003e6aaa49cfe4790848da37681ef1b3e7fc945a765a4228b423aa2b"
177
+ },
178
+ {
179
+ "ep": 600,
180
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep600.pt",
181
+ "sha256": "1f25bb3d58d6e7961fe0ba040dbbc3609f4266c7fadb3876cb6fd425ae409342"
182
+ },
183
+ {
184
+ "ep": 650,
185
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep650.pt",
186
+ "sha256": "ca65208069744e3804e2eb9e48cf263478eac898bd98ae3f7ed10ebf220894c6"
187
+ },
188
+ {
189
+ "ep": 700,
190
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep700.pt",
191
+ "sha256": "b0ade9a8fdcca224d611015a7e49d522449494c4a96d2780bfa8f6afd11af981"
192
+ },
193
+ {
194
+ "ep": 750,
195
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep750.pt",
196
+ "sha256": "28f23c4fc313ed9c22114cf21c84f9ffe42fdb832f5074fe32d114021da09255"
197
+ },
198
+ {
199
+ "ep": 800,
200
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep800.pt",
201
+ "sha256": "c61a6ca8e291d6b91b79d849918456864a67a352be9db199a43305a59a8737e2"
202
+ },
203
+ {
204
+ "ep": 850,
205
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep850.pt",
206
+ "sha256": "0ef3f1b31f053648a213aac470e48b61ed17d96fd4ca3a57c1a50876a4da3213"
207
+ },
208
+ {
209
+ "ep": 900,
210
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep900.pt",
211
+ "sha256": "34e1fa2d7162b16975ac4a563d8834b4490afd06c9c80c1d25b01ae1756edcdb"
212
+ },
213
+ {
214
+ "ep": 950,
215
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep950.pt",
216
+ "sha256": "86de09e5ded1b04518b5dc14eff31316e53b65ae2b5e1e536962591c534e5436"
217
+ },
218
+ {
219
+ "ep": 1000,
220
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1000.pt",
221
+ "sha256": "c39bb132f069a6eae8d6946bb7d791de9ed186e78535778cd9165ead73fd8fb6"
222
+ },
223
+ {
224
+ "ep": 1050,
225
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1050.pt",
226
+ "sha256": "b273de5842aa2e780abf4be0c5c48d98d7ee82a47e455ed20d3fd7af97116da7"
227
+ },
228
+ {
229
+ "ep": 1100,
230
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1100.pt",
231
+ "sha256": "020117fa883e0b6fac3108e2fd2cd5090da75c66bee8631e4548118f577757ca"
232
+ },
233
+ {
234
+ "ep": 1150,
235
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1150.pt",
236
+ "sha256": "cba5fe99dec4ddae23f1f52d8927153fbcd7b7b88a160d87a03293718f4f8e77"
237
+ },
238
+ {
239
+ "ep": 1200,
240
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1200.pt",
241
+ "sha256": "c4f37745b4db8807f4c8c6839d71be4772bcf6e5e760f1c78b66569c10a33071"
242
+ },
243
+ {
244
+ "ep": 1250,
245
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1250.pt",
246
+ "sha256": "91aa6ff2b96efed869b2d595c159923f594688a53b25aef6500b263cfeb6d247"
247
+ },
248
+ {
249
+ "ep": 1300,
250
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1300.pt",
251
+ "sha256": "095e4e619b455ef0361a0bf461c4cc7eaf83d1b279f594ff88beaf4fad36804f"
252
+ },
253
+ {
254
+ "ep": 1350,
255
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1350.pt",
256
+ "sha256": "ee6a4996f0b2afdd46caac3905c6188c7a6939771ab745df175a662ffc470a2b"
257
+ },
258
+ {
259
+ "ep": 1400,
260
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1400.pt",
261
+ "sha256": "a0c2ef8215111bdd7324d1f8e06f3327d181e1dbed8094cf5f1a4341a0e835cd"
262
+ },
263
+ {
264
+ "ep": 1450,
265
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1450.pt",
266
+ "sha256": "55a38946f24469ec90ea6d4ab777790636dc277e42e4889d0520e1ff1aaee534"
267
+ },
268
+ {
269
+ "ep": 1500,
270
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1500.pt",
271
+ "sha256": "8685e395af19c89222a4eac4980db12cd75cb3fe3628db54be3e19ec203c23fe"
272
+ },
273
+ {
274
+ "ep": 1550,
275
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1550.pt",
276
+ "sha256": "a02f87a0d3986548aec82b7bbe9d6ba2fb79316e3f2efb0aa898359a8118ec34"
277
+ },
278
+ {
279
+ "ep": 1600,
280
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1600.pt",
281
+ "sha256": "1d3bde984860d4ed1ad6267c1a0c80edf83dd389aba88c33a0fee8eb4dc5d590"
282
+ },
283
+ {
284
+ "ep": 1650,
285
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1650.pt",
286
+ "sha256": "6c15daa14d292a447f81859b852b7de3dc7e3511f32e7f8b07728f54e9862937"
287
+ },
288
+ {
289
+ "ep": 1700,
290
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1700.pt",
291
+ "sha256": "a8efe775a14fc9e2d7aec9664cc696404c3c3bc2560087b29de021b2d4d374e8"
292
+ },
293
+ {
294
+ "ep": 1800,
295
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1800.pt",
296
+ "sha256": "80683f87b8b3d74962cafb878326c8b8c26081b45e627c31615c383f4f4c6972"
297
+ },
298
+ {
299
+ "ep": 1850,
300
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1850.pt",
301
+ "sha256": "6203fd9e01a8580da3bf179c5c899bd041aae8d4889b096df9bba1cf9917fe65"
302
+ },
303
+ {
304
+ "ep": 1900,
305
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1900.pt",
306
+ "sha256": "014bd83d0e25f354c89aca26af59f4780cd5202b8f55af675b2636e009c864dc"
307
+ },
308
+ {
309
+ "ep": 1950,
310
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep1950.pt",
311
+ "sha256": "4dc7b061d43706cfdbff7efd0fbeff3e47e5738e45b6e4858f7f2fd1bbdfef2b"
312
+ },
313
+ {
314
+ "ep": 2000,
315
+ "ckpt": "ckpt_w9_wcle_i2ce_icetf_g2048_fp_ep2000.pt",
316
+ "sha256": "283d6e93ed7e5c8728c66dbd2d480bf5edbe7c087e3be441a438cc219871455b"
317
+ }
318
+ ]
319
+ }
320
+ },
321
+ "bce_cetf_fixed": {
322
+ "arm": "bce_cetf",
323
+ "fold": null,
324
+ "cap": "",
325
+ "sfx": "",
326
+ "ep": 1350,
327
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1350.pt",
328
+ "sha256": "6c6d0b45bbf73953c8814b2acdd9dd0f3db604662ab58720570f71f3ce9ec552",
329
+ "ranksel": 125.57199015010917,
330
+ "protocol": "fixed split (seed 0)",
331
+ "worker": "Pod/w9_a100_worker.py",
332
+ "selected_on": "the run reads no anchor in training: each of its 40 checkpoints encodes the 4,096-sentence anchor pack of the fixed split (wscan_gal_rev_g4096.npz) and the rank rule picks on that",
333
+ "reads": "Figure 9, Figure A3, Appendix L",
334
+ "extended": "2026-11-08: trained on from 1,000 to 2,000 epochs (weights of ep1000, fresh optimizer, AMP and RNG); rank rule over all 40 checkpoints",
335
+ "trajectory": {
336
+ "why": "Appendix L describes this run's test curve from the selected epoch to the end",
337
+ "files": [
338
+ {
339
+ "ep": 1400,
340
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1400.pt",
341
+ "sha256": "6fa8497122699d00ec1b19c8cd52d6c02a51dea8551ba4c4ad89f77bebfd49b8"
342
+ },
343
+ {
344
+ "ep": 1450,
345
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1450.pt",
346
+ "sha256": "1ab0112d469cb167564fde1081d52c9f3dce1d0f2e07a2f8beabbf1da78f2224"
347
+ },
348
+ {
349
+ "ep": 1500,
350
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1500.pt",
351
+ "sha256": "5806b334a2363dc5ecba3304b24a1eb2f8f3c830164c71f712b839deea644c90"
352
+ },
353
+ {
354
+ "ep": 1550,
355
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1550.pt",
356
+ "sha256": "dfeef1738d00e07ff6d0a982ec1985df9f6e5c800ccc9db57b1e88c9a95be0f6"
357
+ },
358
+ {
359
+ "ep": 1600,
360
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1600.pt",
361
+ "sha256": "4fd7f2d74a59f3be11e994f02f28e18f01bc97839ca92408001030466bc491a0"
362
+ },
363
+ {
364
+ "ep": 1650,
365
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1650.pt",
366
+ "sha256": "d62fa1c8021b893f5e7f29270af3e452aa85e6c3de2e0bcd3b0c3aaa402b4524"
367
+ },
368
+ {
369
+ "ep": 1700,
370
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1700.pt",
371
+ "sha256": "4819efb4080de836e7f4157398132e2f2eed2abe135ba300583e86e3cd3793a9"
372
+ },
373
+ {
374
+ "ep": 1750,
375
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1750.pt",
376
+ "sha256": "cb710c792b7c49fce6a9d15fc65fac4e64a721023cd4d58450e18ca9b7b16a1c"
377
+ },
378
+ {
379
+ "ep": 1800,
380
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1800.pt",
381
+ "sha256": "9d664b78fef3fe8c18afd06b75fd904224da9894e58cd86e9dcfc1dcdf1e29dd"
382
+ },
383
+ {
384
+ "ep": 1850,
385
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1850.pt",
386
+ "sha256": "36f673a0eb425e57ff39c2fcd5774aea6aec2806ee2764a768617160b19cf3f3"
387
+ },
388
+ {
389
+ "ep": 1900,
390
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1900.pt",
391
+ "sha256": "5eb5fe4569975293ddbcad1526363ea6f40e8651031cd2741564d939e88eebc6"
392
+ },
393
+ {
394
+ "ep": 1950,
395
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep1950.pt",
396
+ "sha256": "51b163ca74680536402afcebaada0ec40864a4ded5277de7e8e90a8ede3d55fb"
397
+ },
398
+ {
399
+ "ep": 2000,
400
+ "ckpt": "ckpt_w9_wcle_bce_cetf_fp_ep2000.pt",
401
+ "sha256": "d28075a5ebd76b437fbd06f4bc4f9d2830b8887df79ae1951da71fb1a47ad8b8"
402
+ }
403
+ ]
404
+ }
405
+ },
406
+ "i2bce_icetf_fixed": {
407
+ "arm": "i2bce_icetf",
408
+ "fold": null,
409
+ "cap": "",
410
+ "sfx": "",
411
+ "ep": 1350,
412
+ "ckpt": "ckpt_w9_wcle_i2bce_icetf_fp_ep1350.pt",
413
+ "sha256": "0f3bff568dd396cc473899595237af7a4cdd16e53dcf7e045f2919af9e3eb2b1",
414
+ "ranksel": 125.44088168563394,
415
+ "protocol": "fixed split (seed 0)",
416
+ "worker": "Pod/w9_a100_worker.py",
417
+ "selected_on": "the run reads no anchor in training: each of its 40 checkpoints encodes the 4,096-sentence anchor pack of the fixed split (wscan_gal_rev_g4096.npz) and the rank rule picks on that",
418
+ "reads": "Figure 9, Figure A3, Appendix L",
419
+ "extended": "2026-11-08: trained on from 1,000 to 2,000 epochs (weights of ep1000, fresh optimizer, AMP and RNG); rank rule over all 40 checkpoints"
420
+ },
421
+ "i4uni2_icetf_fixed": {
422
+ "arm": "i4uni2_icetf",
423
+ "fold": null,
424
+ "cap": "",
425
+ "sfx": "",
426
+ "ep": 50,
427
+ "ckpt": "ckpt_w9_wcle_i4uni2_icetf_fp_ep50.pt",
428
+ "sha256": "4e11f8876a7cfb3940cedc237a01fa80cb6d7dba358a0aea19ff7046edc9486c",
429
+ "ranksel": 79.69776134483952,
430
+ "protocol": "fixed split (seed 0)",
431
+ "worker": "Pod/w9_a100_worker.py",
432
+ "selected_on": "the run reads no anchor in training: each of its 40 checkpoints encodes the 4,096-sentence anchor pack of the fixed split (wscan_gal_rev_g4096.npz) and the rank rule picks on that",
433
+ "reads": "Table 3 (a+u @ views), Section 6.2, Appendix L",
434
+ "extended": "2026-11-08: trained on from 1,000 to 2,000 epochs (weights of ep1000, fresh optimizer, AMP and RNG); rank rule over all 40 checkpoints"
435
+ },
436
+ "epd_v25i25c1_cetf_fixed": {
437
+ "arm": "epd_v25i25c1_cetf",
438
+ "fold": null,
439
+ "cap": "",
440
+ "sfx": "",
441
+ "ep": 550,
442
+ "ckpt": "ckpt_w9_wcle_epd_v25i25c1_cetf_fp_ep550.pt",
443
+ "sha256": "67edd01614389dfca04fc65883e22dcf38dfafe8f38a169178604b63165a3f07",
444
+ "ranksel": 81.55530680646169,
445
+ "protocol": "fixed split (seed 0)",
446
+ "worker": "Pod/w9_a100_worker.py",
447
+ "selected_on": "the run reads no anchor in training: each of its 40 checkpoints encodes the 4,096-sentence anchor pack of the fixed split (wscan_gal_rev_g4096.npz) and the rank rule picks on that",
448
+ "reads": "Table A7",
449
+ "extended": "2026-11-08: trained on from 1,000 to 2,000 epochs (weights of ep1000, fresh optimizer, AMP and RNG); rank rule over all 40 checkpoints",
450
+ "weights": "v/i/c = 25/25/1"
451
+ },
452
+ "epd_v20i10c20_cetf_fixed": {
453
+ "arm": "epd_v20i10c20_cetf",
454
+ "fold": null,
455
+ "cap": "",
456
+ "sfx": "",
457
+ "ep": 100,
458
+ "ckpt": "ckpt_w9_wcle_epd_v20i10c20_cetf_fp_ep100.pt",
459
+ "sha256": "fc12e85eb88d9b94bf4a406efbc678ef429a59e1e0c88b4f2d9098725edc4642",
460
+ "ranksel": 106.32286068538902,
461
+ "protocol": "fixed split (seed 0)",
462
+ "worker": "Pod/w9_a100_worker.py",
463
+ "selected_on": "the run reads no anchor in training: each of its 40 checkpoints encodes the 4,096-sentence anchor pack of the fixed split (wscan_gal_rev_g4096.npz) and the rank rule picks on that",
464
+ "reads": "Table A7, Figure 9, Figure A3, Appendix L",
465
+ "extended": "2026-11-08: trained on from 1,000 to 2,000 epochs (weights of ep1000, fresh optimizer, AMP and RNG); rank rule over all 40 checkpoints",
466
+ "weights": "v/i/c = 20/10/20 (recipe weights)"
467
+ },
468
+ "epd_v20i10c15_cetf_fixed": {
469
+ "arm": "epd_v20i10c15_cetf",
470
+ "fold": null,
471
+ "cap": "",
472
+ "sfx": "",
473
+ "ep": 100,
474
+ "ckpt": "ckpt_w9_wcle_epd_v20i10c15_cetf_fp_ep100.pt",
475
+ "sha256": "b87ebec543b7f0adcbb407ecc4fbf3e028d38861325b24f5202301ae977e7e95",
476
+ "ranksel": 106.19608961574562,
477
+ "protocol": "fixed split (seed 0)",
478
+ "worker": "Pod/w9_a100_worker.py",
479
+ "selected_on": "the run reads no anchor in training: each of its 40 checkpoints encodes the 4,096-sentence anchor pack of the fixed split (wscan_gal_rev_g4096.npz) and the rank rule picks on that",
480
+ "reads": "Table A7",
481
+ "extended": "2026-11-08: trained on from 1,000 to 2,000 epochs (weights of ep1000, fresh optimizer, AMP and RNG); rank rule over all 40 checkpoints",
482
+ "weights": "v/i/c = 20/10/15"
483
  }
484
  }
rank_ckpt_plan.json CHANGED
@@ -1488,5 +1488,16 @@
1488
  "init_ckpt": "ckpt_w9cv_wcle_byol_bytf_fold4_gpool_fp_ep300.pt",
1489
  "init_ep": 300,
1490
  "init_sha256": "26bbb5d1abe21c43ed3cc9ab0f2bb3d0ade4a726b8e74f078e17baa06518a216"
 
 
 
 
 
 
 
 
 
 
 
1491
  }
1492
  }
 
1488
  "init_ckpt": "ckpt_w9cv_wcle_byol_bytf_fold4_gpool_fp_ep300.pt",
1489
  "init_ep": 300,
1490
  "init_sha256": "26bbb5d1abe21c43ed3cc9ab0f2bb3d0ade4a726b8e74f078e17baa06518a216"
1491
+ },
1492
+ "dino_dntf_fold0_gpool": {
1493
+ "arm": "dino_dntf",
1494
+ "fold": 0,
1495
+ "cap": "_gpool",
1496
+ "sfx": "",
1497
+ "ep": 1650,
1498
+ "ckpt": "ckpt_w9cv_wcle_dino_dntf_fold0_gpool_fp_ep1650.pt",
1499
+ "sha256": "f2bde75047755c26b713b1d3d31c833bf5cc918dabd742bb7776eecba32d3ffb",
1500
+ "ranksel": 81.45498251422967,
1501
+ "note": "the DINO-style control of Section 4.3 (Table A1), trained on fold 0 only; the file holds the teacher's tower under model, the one read, beside the student and the two prototype heads"
1502
  }
1503
  }