bryanbriney commited on
Commit
813d9ee
·
verified ·
1 Parent(s): 6619c79

Return unused deepclust70 held-out reps to train; drop test_extra; card cleanup

Browse files
README.md CHANGED
@@ -24,8 +24,8 @@ protein language model. Two sources, mixed by rows at training time:
24
  sequences clustered with DIAMOND v2.2.5 `linclust` (`--approx-id 70 --member-cover 90`).
25
  One representative per row, globally shuffled. Re-cut to 1,000,000-row shards by `oplm data reshard`.
26
 
27
- Every number below is copied from `uniref70/summary.json`, `deepclust70/heldout.json` or
28
- `deepclust70/train/summary.json`. The full provenance is in OPLM's
29
  [`docs/data.md`](https://github.com/briney/oplm/blob/main/docs/data.md).
30
 
31
  ## Layout
@@ -42,14 +42,13 @@ uniref70/
42
  test_gte2.parquet # 5,000
43
  deepclust70/
44
  train/summary.json # re-cut record
45
- train/shard-00000.parquet # 5,354 shards, 1,000,000 rows each (the last has 586,520)
46
- heldout.json # held-out screen record
 
47
  eval_singletons.parquet # 5,000 held-out singleton-cluster reps
48
  eval_gte2.parquet # 5,000 held-out reps of 2+-member clusters
49
  test_singletons.parquet # 5,000
50
  test_gte2.parquet # 5,000
51
- test_extra_singletons.parquet # 105,279 (every other screen survivor)
52
- test_extra_gte2.parquet # 79,684
53
  ```
54
 
55
  `manifest.parquet` lists every file with its row count, size and sha256. Check downloads
@@ -98,56 +97,35 @@ Evaluation uses the representatives in both phases.
98
  | `uniref70/eval_gte2.parquet` | 5,000 | 11,119 | 1,973,383 | 4,469,677 |
99
  | `uniref70/test_singletons.parquet` | 5,000 | 5,000 | 1,701,713 | 1,701,713 |
100
  | `uniref70/test_gte2.parquet` | 5,000 | 11,204 | 1,891,172 | 4,295,018 |
101
- | `deepclust70/train/` | 5,353,586,520 | | | |
102
  | `deepclust70/eval_singletons.parquet` | 5,000 | | 486,576 | |
103
  | `deepclust70/eval_gte2.parquet` | 5,000 | | 887,835 | |
104
  | `deepclust70/test_singletons.parquet` | 5,000 | | 488,725 | |
105
  | `deepclust70/test_gte2.parquet` | 5,000 | | 877,340 | |
106
- | `deepclust70/test_extra_singletons.parquet` | 105,279 | | 10,377,799 | |
107
- | `deepclust70/test_extra_gte2.parquet` | 79,684 | | 14,222,233 | |
108
 
109
  uniref70: 121,494,192 UniRef90 sequences in, 75,738,236 clusters out. Members per train
110
  cluster: min 1, median 1, mean 1.604, max 2,370; 60,298,749 train clusters are
111
  singletons.
112
 
113
- ## uniref70 eval and test
114
 
115
- Eval and test clusters are held out of the whole mix, not just uniref70. After the
116
- seeded shuffle, the first 250,000 singleton clusters and the first 250,000 gte2 clusters
117
- are candidates. Every member of every candidate is searched with MMseqs2 against
118
- deepclust70 `train/` and against every member of every other uniref70 cluster:
119
 
120
  ```
121
  --min-seq-id 0.7 -c 0.8 --cov-mode 0 -s 1 --max-seqs 100000 -e 1e6
122
  ```
123
 
124
- A candidate with any hit is dropped. Per stratum, the first 5,000 survivors are eval
125
- and the next 5,000 are test. Candidates not chosen stay in `train/`.
 
 
126
 
127
- | stratum | candidates | survivors | survival |
128
- |---|--:|--:|--:|
129
- | singletons | 250,000 | 64,753 | 25.9% |
130
- | gte2 | 250,000 | 10,832 | 4.3% |
131
-
132
- **Caveat.** The gte2 eval and test sets come from the 4.3% of multi-member clusters
133
- with no ≥ 70%-identity relative anywhere in training. They are a narrower slice than a
134
- random gte2 draw. The build's 5% stop rule tripped on gte2; the maintainers chose to use
135
- the survivors anyway (`--min-survival 0.04`). A bigger pool would give more survivors
136
- at the same rate, because the pool is a uniform draw from the stratum.
137
-
138
- ## deepclust70 eval and test
139
-
140
- deepclust70's original held-out representatives (random draws from singleton clusters
141
- and from clusters with 2+ members; 15,000 eval and 150,000 test per stratum) were
142
- screened with the same MMseqs2 arguments against deepclust70 `train/` and every
143
- uniref70 `train/` member. Per stratum the pool is the old eval file then the old test
144
- file. The first 5,000 survivors are eval, the next 5,000 test, and the rest
145
- `test_extra_*`. The record is `deepclust70/heldout.json`.
146
-
147
- | stratum | candidates | survivors | survival |
148
- |---|--:|--:|--:|
149
- | singletons | 165,000 | 115,279 | 69.9% |
150
- | gte2 | 165,000 | 89,684 | 54.4% |
151
 
152
  ## Rebuild
153
 
@@ -178,6 +156,17 @@ had already been run on its own, with the same `easy-linclust` arguments plus
178
  `--split-memory-limit 80G`, so the run above picked up its output. To rebuild in one
179
  command, add `--mmseqs-extra=--split-memory-limit --mmseqs-extra=80G`.
180
 
 
 
 
 
 
 
 
 
 
 
 
181
  Then write the manifest:
182
 
183
  ```
@@ -189,6 +178,8 @@ oplm data manifest <out>
189
  uniref70 was built with oplm 0.4.1.dev252+ga400dfede. Clustering ran at commit `ea4b33a`
190
  (2026-10-30), the eval screen at `e6f331c` and the write at `a400dfe` (2026-11-01).
191
  The clustering stage's `summary.json` entry was written by hand from that run's record.
 
 
192
 
193
  ## License and attribution
194
 
 
24
  sequences clustered with DIAMOND v2.2.5 `linclust` (`--approx-id 70 --member-cover 90`).
25
  One representative per row, globally shuffled. Re-cut to 1,000,000-row shards by `oplm data reshard`.
26
 
27
+ Every number below is copied from `uniref70/summary.json`, `deepclust70/train/summary.json`
28
+ or `deepclust70/returned.json`. The full provenance is in OPLM's
29
  [`docs/data.md`](https://github.com/briney/oplm/blob/main/docs/data.md).
30
 
31
  ## Layout
 
42
  test_gte2.parquet # 5,000
43
  deepclust70/
44
  train/summary.json # re-cut record
45
+ train/shard-00000.parquet # 5,354 shards, 1,000,000 rows each (the last has 896,515)
46
+ heldout.json # eval/test screen record
47
+ returned.json # record of unused held-out reps returned to train/
48
  eval_singletons.parquet # 5,000 held-out singleton-cluster reps
49
  eval_gte2.parquet # 5,000 held-out reps of 2+-member clusters
50
  test_singletons.parquet # 5,000
51
  test_gte2.parquet # 5,000
 
 
52
  ```
53
 
54
  `manifest.parquet` lists every file with its row count, size and sha256. Check downloads
 
97
  | `uniref70/eval_gte2.parquet` | 5,000 | 11,119 | 1,973,383 | 4,469,677 |
98
  | `uniref70/test_singletons.parquet` | 5,000 | 5,000 | 1,701,713 | 1,701,713 |
99
  | `uniref70/test_gte2.parquet` | 5,000 | 11,204 | 1,891,172 | 4,295,018 |
100
+ | `deepclust70/train/` | 5,353,896,515 | | | |
101
  | `deepclust70/eval_singletons.parquet` | 5,000 | | 486,576 | |
102
  | `deepclust70/eval_gte2.parquet` | 5,000 | | 887,835 | |
103
  | `deepclust70/test_singletons.parquet` | 5,000 | | 488,725 | |
104
  | `deepclust70/test_gte2.parquet` | 5,000 | | 877,340 | |
 
 
105
 
106
  uniref70: 121,494,192 UniRef90 sequences in, 75,738,236 clusters out. Members per train
107
  cluster: min 1, median 1, mean 1.604, max 2,370; 60,298,749 train clusters are
108
  singletons.
109
 
110
+ ## Eval and test
111
 
112
+ Eval and test sets are held out of the whole training mix, not just their own source.
113
+ Each stratum is a random draw of that source's clusters: singletons, or clusters with 2+
114
+ members. Every eval/test sequence is searched with MMseqs2 against deepclust70 `train/`
115
+ and against every member of every uniref70 train cluster:
116
 
117
  ```
118
  --min-seq-id 0.7 -c 0.8 --cov-mode 0 -s 1 --max-seqs 100000 -e 1e6
119
  ```
120
 
121
+ Any hit excludes the example. For uniref70, every member of a held-out cluster is
122
+ searched, not just the representative. Searching every uniref70 member, not just the
123
+ representatives, keeps eval and test independent in phase 2, when training samples
124
+ members.
125
 
126
+ Held-out candidates not used for eval or test are in `train/`, unless they match an
127
+ eval/test sequence under the same search. deepclust70's are shuffled into the last
128
+ `train/` shard.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
129
 
130
  ## Rebuild
131
 
 
156
  `--split-memory-limit 80G`, so the run above picked up its output. To rebuild in one
157
  command, add `--mmseqs-extra=--split-memory-limit --mmseqs-extra=80G`.
158
 
159
+ deepclust70 eval/test, then the unused held-out reps back to `train/`
160
+ (`<held-out dir>` holds deepclust70's original held-out representatives):
161
+
162
+ ```
163
+ oplm data deepclust70-heldout <held-out dir> <out>/deepclust70 \
164
+ --deepclust70 <out>/deepclust70/train --uniref70 <out>/uniref70/train \
165
+ --workdir <work> --threads 32
166
+ oplm data deepclust70-return <held-out dir> <out>/deepclust70 \
167
+ --uniref70 <out>/uniref70 --workdir <work2> --threads 32 --seed 0
168
+ ```
169
+
170
  Then write the manifest:
171
 
172
  ```
 
178
  uniref70 was built with oplm 0.4.1.dev252+ga400dfede. Clustering ran at commit `ea4b33a`
179
  (2026-10-30), the eval screen at `e6f331c` and the write at `a400dfe` (2026-11-01).
180
  The clustering stage's `summary.json` entry was written by hand from that run's record.
181
+ The deepclust70 eval/test screen ran at `ed1d75b` (2026-11-02) and the return of unused
182
+ held-out reps at `fa6d689` (2026-11-04); MMseqs2 18.8cc5c throughout.
183
 
184
  ## License and attribution
185
 
deepclust70/returned.json ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "params": {
3
+ "src": "/home/bryanbriney/data/tmp/deepclust70-heldout/orig",
4
+ "uniref70": "/home/bryanbriney/data/oplm-sequence-data-dev/uniref70",
5
+ "seed": 0,
6
+ "screen_args": [
7
+ "--min-seq-id",
8
+ "0.7",
9
+ "-c",
10
+ "0.8",
11
+ "--cov-mode",
12
+ "0",
13
+ "-s",
14
+ "1",
15
+ "--max-seqs",
16
+ "100000",
17
+ "-e",
18
+ "1e6"
19
+ ]
20
+ },
21
+ "stats": {
22
+ "hits_deepclust70": 0,
23
+ "hits_uniref70": 5,
24
+ "singletons": {
25
+ "unused": 155000,
26
+ "returned": 154998
27
+ },
28
+ "gte2": {
29
+ "unused": 155000,
30
+ "returned": 154997
31
+ },
32
+ "shard": {
33
+ "name": "shard-05353.parquet",
34
+ "rows_before": 586520,
35
+ "rows_after": 896515
36
+ }
37
+ },
38
+ "mmseqs_version": "18.8cc5c",
39
+ "oplm_version": "0.4.1.dev268+gfa6d68911"
40
+ }
deepclust70/test_extra_gte2.parquet DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:7a5b43d73f1f546a9ed09872adaad1e326ea9a3d5214bfe93bf05877a73f8b28
3
- size 8928524
 
 
 
 
deepclust70/test_extra_singletons.parquet DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:12b3e31103c046c03fb3de2992702df8aa986f6726a5e848f4c3e23edb3740fe
3
- size 7113390
 
 
 
 
deepclust70/train/shard-05353.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:43ff038c145f3142b8ec2e69de822d146259fa8075600293b5ff78a72817e241
3
- size 51165989
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0b7c081b66e02f9fb3e1eb35789279d783b1aba2479a24fe020b69fa8deec03e
3
+ size 81311284
manifest.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d78c23dc20f383efc839c3120b95d78858534aa2577015212d145c153eca5978
3
- size 492232
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:08bcbff57dfc849a2963e75a07da38291db488aeb3778de65f33b3b411c0d35f
3
+ size 496617