ift-eval-uk-even / README.md
fassabilf's picture
Eval IFT UK foto asli: gambar + label Gemini, test ep5-20, val 10 titik, held-out 5
c710e60 verified
|
Raw History Blame Contribute Delete
1.64 kB
---
license: creativeml-openrail-m
task_categories: [text-to-image]
tags: [stable-diffusion, fairness, gender-bias, evaluation]
---
# Eval IFT SD 1.5 — foto asli, UK, gender
Gambar hasil generate dan label Gemini untuk arm **foto asli UK** (`fassabilf/sd15-ift-uk-real`,
dataset train `fassabilf/ift-train-uk-real`). 20 okupasi terlatih (sufiks `_uk`), 63 step/epoch.
| split | checkpoint | protokol | n per okupasi |
|---|---|---|---|
| `test_ep*` | ep5, 10, 15, 20 | `pilot_sd15` apa adanya: varian `framing`, `--seed-mode paired --seed-base 0`, `--max-attempts 3`, batch 8 | 100 |
| `val_ep*` | ep1, 2, 3, 4, 6, 8, 10, 12, 16, 20 (dirapatkan di awal) | sama, tapi `--max-attempts 1`, batch 32 | 30 |
| `heldout_ep*` | ep5, 20 | protokol test, 5 okupasi yang **tidak** dilatih: metal_fitter, senior_care_worker, care_worker, crossing_patrol, education_support | 100 |
512 px, 30 step, guidance 7,5, judge **gemini-3.1-pro-preview**. Base per-gambar ada di
mirror `fassabilf/gambarberulangulang` track `pilot_sd15` (`{slug}_uk_framing_r3`).
- `<split>_ep<N>/<okupasi>/predictions.csv` — label per gambar, tanpa perlu unduh zip
- `<split>_ep<N>/<okupasi>/manifest.csv`, `gemini_judge_meta.json` — setelan judge persis
- `zips/<split>_ep<N>__<okupasi>.zip` — gambarnya
- `analysis/` — CSV bootstrap, paired Δ vs base, lantai MAE, fixed-t val loss
Angka utama (paired ΔMAE_real vs base, cluster bootstrap 95 %): ep5 −0,044 [−0,090, −0,003],
ep10 −0,027 [−0,097, +0,045], ep15 −0,055 [−0,114, −0,002], ep20 −0,057 [−0,099, −0,018].
n=30 per okupasi di val → SD sampling ≈ 0,02; kurvanya bergerigi.