# Development history

This is the previous model card, kept verbatim.

---
license: cc-by-4.0
language:
- en
task_categories:
- image-segmentation
tags:
- optical-gas-imaging
- audit
- weak-masks
configs:
- config_name: recording_manifest
  data_files:
  - split: development_train
    path: recording_manifest/development_train.csv
  - split: development_tune
    path: recording_manifest/development_tune.csv
  - split: development_compare
    path: recording_manifest/development_compare.csv
- config_name: compare_frames
  data_files:
  - split: development_compare
    path: compare_frames.csv
- config_name: candidates
  data_files:
  - split: development_tune
    path: candidates.csv
---

# Controlled Diet recording/mask audit

**137 audit rows describe 42 selected source images in nine recording IDs.**
They are not 137 images, independent releases, chemical measurements or gas-free
sessions. This native CSV dataset transfers the reviewed frozen CPU JSON records;
it adds no source pixel access, scoring, fitting or model selection.

| Configuration | Development roles | Rows |
|---|---|---:|
| recording_manifest | development_train / development_tune / development_compare | 18 / 12 / 12 = 42 |
| compare_frames | development_compare | 12 filenames × 7 frozen arms = 84 |
| candidates | development_tune | All 11 candidates, including the six learned fits |

`recording_manifest` exposes original filenames, recording IDs, source paths/full
LFS SHA256/bytes, dietary class IDs and publisher foreground counts. Original
publisher splits are descriptive; the study assigns complete recording IDs to
disjoint development roles. `recording_selection.json` contains the same minimal
42 row records. Source dietary IDs are not per-frame concentration measurements.
All source foreground counts are positive; physical-release independence is unknown.

`compare_frames` exposes native 640×480 integer TP/FP/FN/TN and original nullable
IoU/Dice/precision/recall/`background_pixel_overprediction_ratio`. Counts refer to
publisher-generated nonzero masks. Undefined metrics are empty CSV cells, not zero.
The background-pixel ratio FP/(FP+TN) is weak-reference pixel overprediction; it is
never a gas false-alarm rate, gas-absence score or count of negative recording sessions.
Primary aggregation is equal-weight recording means of frame IoU, as defined in
`protocol.json`; these rows are per-frame evidence, not a new pooled selection rule.

`candidates` retains completed outcomes, frozen selection flags, L2/step counts,
objective endpoints/trace counts and tune recording-macro IoU. Baselines' undefined
training objectives/L2 are empty cells. Comparison did not retune/refit/promote.
The spatial prior comparison IoU is 0.138078, tiny MLP 0.026416, logistic 0.022555:
both learned models lose. No industrial pipeline integration or field/chemical/device
qualification is provided. All roles remain exploratory `development_*` roles.

## Source, attribution and modifications

Source: [Controlled Diet (CD) Dataset for Methane Plume Detection](https://huggingface.co/datasets/baselab/controlled-diet-methane-dataset),
by Mohamed G. Embaby, Toqi Tahamid Sarker, Amer AbuGhazaleh and Khaled R. Ahmed,
Southern Illinois University Carbondale. Exact revision `9475f96c32dd4764471e4358ca6659b0dd348ecd`.
[Pinned source license](https://huggingface.co/datasets/baselab/controlled-diet-methane-dataset/blob/9475f96c32dd4764471e4358ca6659b0dd348ecd/LICENSE)
declares CC BY 4.0; a source metadata JSON separately declares CC0. The unresolved
discrepancy is retained; this audit elects CC BY 4.0 and makes no public-domain claim.
See [ATTRIBUTION.md](ATTRIBUTION.md), [LICENSE](LICENSE) and
[CC BY 4.0 terms](https://creativecommons.org/licenses/by/4.0/). No endorsement is implied.

Changes: prospectively selected 42 pairs; recording-disjoint development roles;
nonzero binary mask occupancy; fractional training targets and native-resolution
confusion counts; CSV transfer of the frozen source IDs and CPU results. Filenames,
source hashes, bytes and publisher counts are retained. No copied article or image
payload is in this audit repository. The related model's original cloud archive
retains the qualified PNG subset with its source notices and separate dependency grants.

Selection SHA256: `467c1c8595dc1a23365f80038771c86d35542640e77c6b9924640cb8655a2adb`.
Complete evidence/model/source: https://huggingface.co/sankalpsthakur/controlled-diet-ogi-mask-reference
Read-only viewer target: https://huggingface.co/spaces/sankalpsthakur/cd-ogi-mask-reference-audit
Anonymous fixed-revision file readback and all native audit tables now pass. The [publication record](https://huggingface.co/sankalpsthakur/controlled-diet-ogi-mask-reference/blob/main/PUBLICATION.md) identifies exact content and observed Viewer revisions; generic Viewer URLs remain cached selectors.