Updated the doc about fighting bias in alt text.
tldr; using GPT 4 omnium to generate synthetic alt texts following our guidelines gives pretty good results.
I've ran in against Flickr30k => https://huggingface.co/datasets/Mozilla/flickr30k-transformed-captions-gpt4o so we can train again on a "cleaner" dataset.
https://github.com/mozilla/distilvit/blob/main/docs/fighting_bias.md
Tarek Ziadé :firefox: /
npub1jv…mnkp0
2024-07-05 07:34:25
Author Public Key
npub1jvdce7j268umhqcse6pq5y956suw35af7j9gx0rn6zxpsec582aqamnkp0Published at
2024-07-05 07:34:25Event JSON
{
"id": "7068026d31c5e70699126f85b8dc1108ce9b041ee0a5642c389ad68e1b29ad0d",
"pubkey": "931b8cfa4ad1f9bb8310ce820a10b4d438e8d3a9f48a833c73d08c1867143aba",
"created_at": 1720164865,
"kind": 1,
"tags": [
[
"proxy",
"https://framapiaf.org/@tarek/112732724651339426",
"web"
],
[
"proxy",
"https://framapiaf.org/users/tarek/statuses/112732724651339426",
"activitypub"
],
[
"L",
"pink.momostr"
],
[
"l",
"pink.momostr.activitypub:https://framapiaf.org/users/tarek/statuses/112732724651339426",
"pink.momostr"
],
[
"expiration",
"1722780094"
]
],
"content": "Updated the doc about fighting bias in alt text.\n\ntldr; using GPT 4 omnium to generate synthetic alt texts following our guidelines gives pretty good results.\n\nI've ran in against Flickr30k =\u003e https://huggingface.co/datasets/Mozilla/flickr30k-transformed-captions-gpt4o so we can train again on a \"cleaner\" dataset.\n\nhttps://github.com/mozilla/distilvit/blob/main/docs/fighting_bias.md",
"sig": "7d4359563c3bad8fb095dd5ec3fe57d8a2560f9c728894803288025d43360f06fa7346e71bff58b096cfa76b3a3567e154ca3584852e672f9b8f75c8d5bd906c"
}