YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Panoptes

Ink detection checkpoints from the ScrollAgent pipeline studies of October 2026, kept public because the article uses them as evidence of what worked and what did not: training on three scrolls at full and at rule-stopped steps, the label pair (our corrected labels against the Reader v2 labels), the learned label offset, and the Reader V3 stage. thin80 is the core AUC of the private bench files named in each row, on unit w060, and minus_rv2_thin80 its difference to the published Reader v2 on the same unit. Each model has a card under cards/ and its checkpoints under ///; internal tags and folder paths are unchanged.

Licence: Weights trained on Vesuvius Challenge data; released for non-commercial research use with attribution (CC BY-NC 4.0), subject to the data terms at scrollprize.org.

public name folder stage steps reached / planned what it shows in the article
villa-hybrid-exact3-tri-s42-s1-v2cb gpu1g9e3v2c/chain-tri-s42 s1 78125 / 78125 The s1 stage (organisers' hybrid recipe, 78125 fixed steps) that the s2 row below starts from.
readerv2-recipe-on-villa-hybrid-exact3-tri-s42-from-s1-ckpt078125-s2 gpu1g9e3v2c/chain-tri-s42 s2 40000 / 40000 Three training scrolls at full fixed steps (s1 78125, s2 40000), the reference arm. CBF: thin80 0.8193892954935759, minus_rv2_thin80 0.0008330595925813977 (v2crc-results.csv line 10, pass 2, unit w060, 116 keV, raw labels).
villa-hybrid-exact3-tri-s43-s1-s1x gpu1g9e3s1x/chain-tri-s43 s1 54000 / 78125 The s1 stage of the second seed, ended by the declared gain rule; its ledger row (2026-10-20T16:13:11Z pipeline-as-system-gpu1g9e3s1x-autostop) reads: «chain-tri-s43: stage s1 stop confirmed by the rule (two evaluations without a 0.01 % gain, at or after the minimum 46875) at the validation of step 56250 (min 46875); KEPT s1/ckpt_050001.pth sha256 0ca33454b83f (val_loss 0.575180, pulled and pinned); stage-stop-s1 written on the volume»
readerv2-recipe-on-villa-hybrid-exact3-tri-s43-from-s1-ckpt050001-s2 gpu1g9e3s1x/chain-tri-s43 s2 40000 / 40000 Second seed, s2 at 40000 steps after the rule-stopped s1. T43F: thin80 0.8251498967499971, minus_rv2_thin80 0.006593660849002592 (t43r-results.csv line 6, pass none, unit w060, 116 keV, raw labels).
villa-hybrid-exact3-tri-s42-s1-gpu1g9e3 gpu1g9e3/chain-tri-s42 s1 47000 / 78125 Only ckpt_025001.pth is kept from this stage: the s1 checkpoint the s2x row below starts from (sha256 804838c9d3a67cce52b44747e7f10e6bae734295c9b3391dacedfc8b69cc956f, recorded in the s2 start.json of the job's first run).
readerv2-recipe-on-villa-hybrid-exact3-tri-s42-s2 gpu1g9e3s2x/chain-tri-s42 s2 40000 / 40000 The label pair, first half: exact3 corrected labels, s2 completed to 40000 steps. S42F: thin80 0.8092220575258902, minus_rv2_thin80 -0.009334178375104307 (v2crc-results.csv line 6, pass 2, unit w060, 116 keV, raw labels); minus S42UF on the same unit: -0.004374.
villa-hybrid-readerv2-labels-tri-s42-s1-gpu1g9e3 gpu1g9e3/chain-tri-s42-u s1 46875 / 78125 Only ckpt_031251.pth is kept from this stage: the s1 checkpoint the s2x row below starts from (sha256 65d24daff26f7943f22e4dcaf56332d97baaa509646345475d6c1728e17b3299, recorded in the s2 start.json of the job's first run).
readerv2-recipe-on-villa-hybrid-readerv2-labels-tri-s42-s2 gpu1g9e3s2x/chain-tri-s42-u s2 40000 / 40000 The label pair, second half: the same chain and seed on the Reader v2 labels, without our correction. S42UF: thin80 0.8135958253386826, minus_rv2_thin80 -0.00496041056231189 (v2crc-results.csv line 8, pass 2, unit w060, 116 keV, raw labels).
villa-hybrid-exact3-tri-s42-s1-v2lob gpu1g9e3v2lob/chain-tri-s42 s1 78125 / 78125 The s1 stage of the learned-offset twin of gpu1g9e3v2c, with the same fixed steps and one training change, a learned per segment label offset.
readerv2-recipe-on-villa-hybrid-exact3-tri-learned-offset-s42-from-s1-ckpt078125-s2 gpu1g9e3v2lob/chain-tri-s42 s2 40000 / 40000 The learned-offset negative result. LF: thin80 0.8150576873729606, minus_rv2_thin80 -0.0034985485280338846 (v2crc-results.csv line 14, pass 2, unit w060, 116 keV, raw labels); minus CBF, its twin without the offset, on the same unit: -0.004332.
v3-on-readerv2-recipe-on-villa-hybrid-exact3-tri-p256-s42-from-s2-ckpt040000-s3 gpu1g9e3v3b/chain-tri-s42 s3 30000 / 30000 The Reader V3 stage (17 x 256 x 256 patches, 30000 fixed steps) on top of the three-scroll s2 of gpu1g9e3v2c, read at 128 px crop. B3E: thin80 0.816904225003309, minus_rv2_thin80 -0.0016520108976855186 (v3br2-results.csv line 17, pass 2, unit w060, 116 keV, raw labels).

The other models of these studies were recipe studies, partial or failed runs; they were removed from the public archive.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support