Skip to content

Commit 3d53c3f

Browse files
cursoragentznsoftm
andcommitted
Record a2_cn_crops parts 006-007 and the 325k pretrain eval.
The backup range moved forward, and validation exact match rose to 4/344. Co-authored-by: Daniel <znsoft@163.com>
1 parent 8c33a9e commit 3d53c3f

1 file changed

Lines changed: 5 additions & 4 deletions

File tree

‎docs/HANDOVER.md‎

Lines changed: 5 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,6 @@
11
# RapidStamp ring-text handover
22

3-
Updated 2026-10-23T06:12:41Z. This file is the operational handover. `docs/data.md` and `docs/research.md` are the Chinese measured notes; `docs/sdk.md` is the published detection plus class-routed CTC / PP-OCRv5 SDK. Those three are not this recognizer.
3+
Updated 2026-10-23T06:15:32Z. This file is the operational handover. `docs/data.md` and `docs/research.md` are the Chinese measured notes; `docs/sdk.md` is the published detection plus class-routed CTC / PP-OCRv5 SDK. Those three are not this recognizer.
44

55
Update this file after every meaningful step, measured result, or change to data or models. Commit it on `main`, then copy `docs/*.md` to `/home/rapidstamp_data/docs/` on the training machine. Do not put passwords, the SSH password, or Aliyun APPCODEs in this file or in git.
66

@@ -54,7 +54,7 @@ A second copy of the collection box is arriving as five plain tar archives, spli
5454
| Archive | Parts | Bytes | Files | Received |
5555
| --- | --- | --- | --- | --- |
5656
| `a1_meta.tar` | 20 | 442,204,160 | 1,124 | complete. Tar sha256 `f2556116d5aa80cdd48af47887a72a40274fff0d398de67620c711b0da25f0dc`. Extracted. |
57-
| `a2_cn_crops.tar` | 60 | 1,361,745,920 | 26,265 | 000–005, sha256 checked on the server |
57+
| `a2_cn_crops.tar` | 60 | 1,361,745,920 | 26,265 | 000–007, sha256 checked on the server |
5858
| `a3_foreign_crops.tar` | 26 | 591,646,720 | 4,787 | none |
5959
| `a4_cn_rejected.tar` | 13 | 277,166,080 | 4,350 | none |
6060
| `a5_foreign_rejected.tar` | 36 | 813,629,440 | 3,637 | none |
@@ -123,7 +123,7 @@ For each query it scrapes Bing, Baidu, and 360 image search, crops seals with th
123123
| ViT-Tiny direct reader, epoch 33 of 40 | 9/344, 0.31392 | 0/63, 0.18681 (verified 0/57, 0.19851; pending 0/6, 0.07564) | Best by val exact. Predictions are mostly generic shells. One val near-miss: 吉南省人力资源和社会保障厅 vs 吉林省人力资源和社会保障行. |
124124
| PARSeq ViT-S pretrain, in progress | see below | see below | Not a result yet. |
125125

126-
Pretraining snapshot, job still running, log tail at 2026-10-23T06:12:41Z (PID 52706): seen 308,224, loss 6.445, 110.6 img/s. Eval so far (1−NED; gold exact is still 0; val exact is 1/344 at 225,024 and 250,016 and 275,008, then 3/344 at 300,000):
126+
Pretraining snapshot, job still running, log tail at 2026-10-23T06:15:32Z (PID 52706): seen 326,816, loss 6.321, 110.6 img/s. Eval so far (1−NED; gold exact is still 0; val exact is 1/344 at 225,024, 250,016, and 275,008, then 3/344 at 300,000, then 4/344 at 325,024):
127127

128128
| Seen | Val 1−NED (n=344) | Gold 1−NED (n=63) | Verified (57) | Pending (6) |
129129
| --- | --- | --- | --- | --- |
@@ -139,8 +139,9 @@ Pretraining snapshot, job still running, log tail at 2026-10-23T06:12:41Z (PID 5
139139
| 250,016 | 0.36724 | 0.25206 | 0.27077 | 0.07423 |
140140
| 275,008 | 0.37067 | 0.26332 | 0.27761 | 0.12756 |
141141
| 300,000 | 0.42490 | 0.27651 | 0.28944 | 0.15370 |
142+
| 325,024 | 0.47119 | 0.28823 | 0.29877 | 0.18808 |
142143

143-
`eval.json` on disk matches the 300,000 row. Val exact is 3/344. Gold exact is still 0/63. Val and gold 1−NED are still rising after the dip at 75,008 and 100,000. Predictions are still mostly generic shells. Loss started near 10 (two-pass CE; `log(6627) ≈ 8.8`) and is falling. These rows are early. The zero-shot number is the eval written when the 2,000,000-sample stage ends.
144+
`eval.json` on disk matches the 325,024 row. Val exact is 4/344. Gold exact is still 0/63. Val and gold 1−NED are still rising after the dip at 75,008 and 100,000. Predictions are still mostly generic shells. Loss started near 10 (two-pass CE; `log(6627) ≈ 8.8`) and is falling. These rows are early. The zero-shot number is the eval written when the 2,000,000-sample stage ends.
144145

145146
## Checkpoints and ONNX
146147

0 commit comments

Comments
 (0)