@percyliang
Normally replay old data reduces forgetting, but it actually helps you learn on new data too! We finally put this paper out on arxiv, but had it up as a Marin GitHub issue ~1 year ago: https://t.co/MNevf6XjvC
Viewing enriched Twitter post
Normally replay old data reduces forgetting, but it actually helps you learn on new data too! We finally put this paper out on arxiv, but had it up as a Marin GitHub issue ~1 year ago: https://t.co/MNevf6XjvC
{
"media": [
{
"type": "photo",
"url": "https://crmoxkoizveukayfjuyo.supabase.co/storage/v1/object/public/media/posts/2030084101559271490/media_0.jpg?",
"filename": "media_0.jpg"
}
],
"processed_at": "2026-03-07T14:27:17.277285",
"pipeline_version": "2.0"
} {
"type": "tweet",
"id": "2030084101559271490",
"url": "https://x.com/percyliang/status/2030084101559271490",
"twitterUrl": "https://twitter.com/percyliang/status/2030084101559271490",
"text": "Normally replay old data reduces forgetting, but it actually helps you learn on new data too!\n\nWe finally put this paper out on arxiv, but had it up as a Marin GitHub issue ~1 year ago:\nhttps://t.co/MNevf6XjvC",
"source": "Twitter for iPhone",
"retweetCount": 14,
"replyCount": 6,
"likeCount": 143,
"quoteCount": 0,
"viewCount": 15403,
"createdAt": "Sat Mar 07 00:52:21 +0000 2026",
"lang": "en",
"bookmarkCount": 87,
"isReply": false,
"inReplyToId": null,
"conversationId": "2030084101559271490",
"displayTextRange": [
0,
209
],
"inReplyToUserId": null,
"inReplyToUsername": null,
"author": {
"type": "user",
"userName": "percyliang",
"url": "https://x.com/percyliang",
"twitterUrl": "https://twitter.com/percyliang",
"id": "86481377",
"name": "Percy Liang",
"isVerified": false,
"isBlueVerified": true,
"verifiedType": null,
"profilePicture": "https://pbs.twimg.com/profile_images/1319828397699129346/wjBcgUH0_normal.jpg",
"coverPicture": "",
"description": "",
"location": "Stanford, CA",
"followers": 97005,
"following": 424,
"status": "",
"canDm": false,
"canMediaTag": true,
"createdAt": "Sat Oct 31 07:26:37 +0000 2009",
"entities": {
"description": {
"urls": []
},
"url": {}
},
"fastFollowersCount": 0,
"favouritesCount": 2582,
"hasCustomTimelines": true,
"isTranslator": false,
"mediaCount": 105,
"statusesCount": 1258,
"withheldInCountries": [],
"affiliatesHighlightedLabel": {},
"possiblySensitive": false,
"pinnedTweetIds": [
"1924527490351169964"
],
"profile_bio": {
"description": "professor of computer science @Stanford @stanfordnlp, co-founder of @togethercompute, creator of https://t.co/7R5THVogW2, co-founder of @simile_ai, pianist",
"entities": {
"description": {
"hashtags": [],
"symbols": [],
"urls": [
{
"display_url": "marin.community",
"expanded_url": "http://marin.community",
"indices": [
97,
120
],
"url": "https://t.co/7R5THVogW2"
}
],
"user_mentions": [
{
"id_str": "0",
"indices": [
30,
39
],
"name": "",
"screen_name": "Stanford"
},
{
"id_str": "0",
"indices": [
40,
52
],
"name": "",
"screen_name": "stanfordnlp"
},
{
"id_str": "0",
"indices": [
68,
84
],
"name": "",
"screen_name": "togethercompute"
},
{
"id_str": "0",
"indices": [
136,
146
],
"name": "",
"screen_name": "simile_ai"
}
]
},
"url": {
"urls": [
{
"display_url": "cs.stanford.edu/~pliang/",
"expanded_url": "https://cs.stanford.edu/~pliang/",
"indices": [
0,
23
],
"url": "https://t.co/eDtsxiCvFO"
}
]
}
}
},
"isAutomated": false,
"automatedBy": null
},
"extendedEntities": {},
"card": {
"binding_values": [
{
"key": "description",
"value": {
"string_value": "Description Interested in how to properly order pretraining data for fixed available data. To start, we can consider how you would want to do curriculum learning for transfer: if you are given a fe..."
}
},
{
"key": "domain",
"value": {
"string_value": "github.com"
}
},
{
"key": "site",
"value": {
"scribe_key": "publisher_id",
"user_value": {
"id_str": "13334762",
"path": []
}
}
},
{
"key": "photo_image_full_size_alt_text",
"value": {
"string_value": "Description Interested in how to properly order pretraining data for fixed available data. To start, we can consider how you would want to do curriculum learning for transfer: if you are given a fe..."
}
},
{
"key": "vanity_url",
"value": {
"scribe_key": "vanity_url",
"string_value": "github.com"
}
},
{
"key": "summary_photo_image_alt_text",
"value": {
"string_value": "Description Interested in how to properly order pretraining data for fixed available data. To start, we can consider how you would want to do curriculum learning for transfer: if you are given a fe..."
}
},
{
"key": "title",
"value": {
"string_value": "Studying how data ordering can improve pretraining · Issue #702 · marin-community/marin"
}
},
{
"key": "card_url",
"value": {
"scribe_key": "card_url",
"string_value": "https://t.co/MNevf6XjvC"
}
}
],
"card_platform": {
"platform": {
"audience": {
"name": "production"
},
"device": {
"name": "iPhone",
"version": "13"
}
}
},
"name": "summary_large_image",
"url": "https://t.co/MNevf6XjvC",
"user_refs_results": [
{
"rest_id": "13334762",
"result": {
"__typename": "User",
"action_counts": {
"favorites_count": 8640
},
"avatar": {
"image_url": "https://pbs.twimg.com/profile_images/1633247750010830848/8zfRrYjA_normal.png"
},
"banner": {
"image_url": "https://pbs.twimg.com/profile_banners/13334762/1765308302"
},
"core": {
"created_at": "Mon Feb 11 04:41:50 +0000 2008",
"name": "GitHub",
"screen_name": "github"
},
"dm_permissions": {
"can_dm": false
},
"exclusive_tweet_following": false,
"follow_request_sent": false,
"identity_profile_labels_highlighted_label": {},
"location": {
"location": "San Francisco, CA"
},
"media_permissions": {
"can_media_tag": true
},
"notifications_settings": {
"notifications_enabled": false
},
"pinned_items": {
"tweet_ids_str": [
"2019093909981257849"
]
},
"privacy": {
"protected": false,
"suspended": false
},
"private_super_following": false,
"profile_bio": {
"description": "The AI-powered developer platform to build, scale, and deliver secure software.",
"entities": {
"description": {
"hashtags": [],
"symbols": [],
"urls": [],
"user_mentions": []
},
"url": {
"urls": [
{
"display_url": "github.com",
"expanded_url": "http://github.com",
"indices": [
0,
23
],
"url": "https://t.co/bbJgfyzKzp"
}
]
}
}
},
"profile_image_shape": "Square",
"profile_metadata": {
"profile_interstitial_type": "",
"profile_link_color": "981CEB"
},
"profile_translation": {
"translator_type_enum": "None"
},
"properties": {
"has_extended_profile": true
},
"relationship_counts": {
"followers": 2619287,
"following": 333
},
"relationship_perspectives": {
"blocked_by": false,
"blocking": false,
"followed_by": false,
"following": false,
"live_following": false,
"muting": false
},
"rest_id": "13334762",
"smart_blocked_by": false,
"smart_blocking": false,
"super_follow_eligible": false,
"super_followed_by": false,
"super_following": false,
"tweet_counts": {
"media_tweets": 2914,
"tweets": 10395
},
"verification": {
"is_blue_verified": true,
"verified": false,
"verified_type": "Business"
},
"website": {
"url": "https://t.co/bbJgfyzKzp"
}
}
}
]
},
"place": {},
"entities": {
"hashtags": [],
"symbols": [],
"timestamps": [],
"urls": [
{
"display_url": "github.com/marin-communit…",
"expanded_url": "https://github.com/marin-community/marin/issues/702",
"indices": [
186,
209
],
"url": "https://t.co/MNevf6XjvC"
}
],
"user_mentions": []
},
"quoted_tweet": {
"type": "tweet",
"id": "2029983689988542742",
"url": "https://x.com/kothasuhas/status/2029983689988542742",
"twitterUrl": "https://twitter.com/kothasuhas/status/2029983689988542742",
"text": "to improve fine-tuning data efficiency, replay generic pre-training data\n\nnot only does this reduce forgetting, it actually improves performance on the fine-tuning domain! especially when fine-tuning data is scarce in pre-training (w/ @percyliang) https://t.co/ClGPAUlPqQ",
"source": "Twitter for iPhone",
"retweetCount": 46,
"replyCount": 10,
"likeCount": 395,
"quoteCount": 2,
"viewCount": 35764,
"createdAt": "Fri Mar 06 18:13:21 +0000 2026",
"lang": "en",
"bookmarkCount": 283,
"isReply": false,
"inReplyToId": null,
"conversationId": "2029983689988542742",
"displayTextRange": [
0,
247
],
"inReplyToUserId": null,
"inReplyToUsername": null,
"author": {
"type": "user",
"userName": "kothasuhas",
"url": "https://x.com/kothasuhas",
"twitterUrl": "https://twitter.com/kothasuhas",
"id": "1266445726491410432",
"name": "Suhas Kotha",
"isVerified": false,
"isBlueVerified": false,
"verifiedType": null,
"profilePicture": "https://pbs.twimg.com/profile_images/1694095997184376833/0nuYqdDE_normal.jpg",
"coverPicture": "https://pbs.twimg.com/profile_banners/1266445726491410432/1695070580",
"description": "",
"location": "",
"followers": 725,
"following": 274,
"status": "",
"canDm": true,
"canMediaTag": true,
"createdAt": "Fri May 29 19:06:10 +0000 2020",
"entities": {
"description": {
"urls": []
},
"url": {}
},
"fastFollowersCount": 0,
"favouritesCount": 546,
"hasCustomTimelines": true,
"isTranslator": false,
"mediaCount": 27,
"statusesCount": 127,
"withheldInCountries": [],
"affiliatesHighlightedLabel": {},
"possiblySensitive": false,
"pinnedTweetIds": [
"1969094885518053869"
],
"profile_bio": {
"description": "cs phd @ stanford",
"entities": {
"description": {
"hashtags": [],
"symbols": [],
"urls": [],
"user_mentions": []
},
"url": {
"urls": [
{
"display_url": "kothasuhas.github.io",
"expanded_url": "http://kothasuhas.github.io/",
"indices": [
0,
23
],
"url": "https://t.co/kT6iprmMcx"
}
]
}
}
},
"isAutomated": false,
"automatedBy": null
},
"extendedEntities": {
"media": [
{
"allow_download_status": {
"allow_download": true
},
"display_url": "pic.twitter.com/ClGPAUlPqQ",
"expanded_url": "https://twitter.com/kothasuhas/status/2029983689988542742/photo/1",
"ext_media_availability": {
"status": "Available"
},
"features": {
"all": {
"tags": [
{
"name": "Percy Liang",
"screen_name": "percyliang",
"type": "user",
"user_id": "86481377"
}
]
},
"large": {
"faces": []
},
"orig": {
"faces": []
}
},
"id_str": "2029982860254588928",
"indices": [
248,
271
],
"media_key": "3_2029982860254588928",
"media_results": {
"id": "QXBpTWVkaWFSZXN1bHRzOgwAAQoAARwr8qf1mzAACgACHCvzaSWakRYAAA==",
"result": {
"__typename": "ApiMedia",
"id": "QXBpTWVkaWE6DAABCgABHCvyp/WbMAAKAAIcK/NpJZqRFgAA",
"media_key": "3_2029982860254588928"
}
},
"media_url_https": "https://pbs.twimg.com/media/HCvyp_WbMAAw6Rr.jpg",
"original_info": {
"focus_rects": [
{
"h": 819,
"w": 1462,
"x": 0,
"y": 0
},
{
"h": 958,
"w": 958,
"x": 287,
"y": 0
},
{
"h": 958,
"w": 840,
"x": 346,
"y": 0
},
{
"h": 958,
"w": 479,
"x": 527,
"y": 0
},
{
"h": 958,
"w": 1462,
"x": 0,
"y": 0
}
],
"height": 958,
"width": 1462
},
"sizes": {
"large": {
"h": 958,
"w": 1462
}
},
"type": "photo",
"url": "https://t.co/ClGPAUlPqQ"
}
]
},
"card": null,
"place": {},
"entities": {
"hashtags": [],
"symbols": [],
"timestamps": [],
"urls": [],
"user_mentions": [
{
"id_str": "86481377",
"indices": [
235,
246
],
"name": "Percy Liang",
"screen_name": "percyliang"
}
]
},
"quoted_tweet": null,
"retweeted_tweet": null,
"isLimitedReply": false,
"article": null
},
"retweeted_tweet": null,
"isLimitedReply": false,
"article": null
}