Datasetpublic

fassabilf/ift-eval-us-dpo

Gambar eval — arm DPO (foto asli & sintetis) dan SFT even, US, gender, SD 1.5 Semua gambar di sini dirender model (bukan foto asli), plus label judge gemini-3.1-pro-preview per gambar. Protokol pilot_sd15: 14 okupasi, n=100/okupasi (test), 30 prompt/okupasi (val). folder model checkpoint dpo_sd15/test_ep* fassabilf/sd15-dpo-us-sd15 ep5/10/15/20 dpo_real/test_ep* fassabilf/sd15-dpo-us-real ep5/10/15/20 dpo_real/val_ep* fassabilf/sd15-dpo-us-real ep2..20 tiap 2… See the full description on the dataset page: https://huggingface.co/datasets/fassabilf/ift-eval-us-dpo.

sourceHugging Facecreativeml-openrail-mupdated 12d agoView on Hugging Face

0likes260downloads

Dataset Card

Gambar eval — arm DPO (foto asli & sintetis) dan SFT even, US, gender, SD 1.5

Semua gambar di sini dirender model (bukan foto asli), plus label judge gemini-3.1-pro-preview per gambar. Protokol pilot_sd15: 14 okupasi, n=100/okupasi (test), 30 prompt/okupasi (val).

folder	model	checkpoint
`dpo_sd15/test_ep*`	`fassabilf/sd15-dpo-us-sd15`	ep5/10/15/20
`dpo_real/test_ep*`	`fassabilf/sd15-dpo-us-real`	ep5/10/15/20
`dpo_real/val_ep*`	`fassabilf/sd15-dpo-us-real`	ep2..20 tiap 2
`sft_even/test_ep*`	`fassabilf/sd15-ift-us-even`	ep5/10/15

Tiap <folder>/<okupasi>/ berisi predictions.csv (label judge), manifest.csv, gemini_judge_meta.json — terbuka, bisa dibaca tanpa unduh gambar. Gambarnya ada di zips/<arm>__<folder>.zip (ZIPSTORED); unzip ke `outputs/ift{test,val}us<arm>_ep<N>/ di repo GambarBerulangUlang` untuk menjalankan ulang analisis tanpa render/judge ulang.

eval_csv/ = CSV agregat outputs/ift_eval_us_*.csv. workspace/ = catatan run dan log pod (PROGRESS.md, skrip peluncur, log training/eval/upload).

Analisis & laporan: GambarBerulangUlang/analysis/{dpo_real,dpo_sd15}/.