File size: 3,743 Bytes
fbe8d49 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 | """Mix definition: million words per source for the 10M / 100M / 1000M versions."""
SIZES = {"10M": 0, "100M": 1, "1000M": 2}
# key: (genre, (10M, 100M, 1000M) quota in million words)
MIX = {
"tinystories": ("narrative", (1.5, 8, 40)),
"simplestories": ("narrative", (2.0, 8, 60)),
"cosmo_stories": ("narrative", (0, 5, 50)),
"tinydialogues": ("dialogue", (1.5, 7, 29)),
"soda": ("dialogue", (1, 9, 75)),
"ultrachat": ("dialogue", (0, 4, 61)),
"smoltalk": ("dialogue", (0, 1, 25)),
"tinyhelen": ("expository", (1.5, 10, 50)),
"simplewiki": ("expository", (1, 10, 30)),
"enwiki_intro": ("expository", (0, 0, 50)),
"cosmo_v2": ("expository", (1, 12, 180)),
"tiny_textbooks": ("expository", (0, 4, 40)), # uses the synthetic `textbook` column
"cosmo_v01": ("expository", (0.5, 3, 30)),
"fineweb_edu": ("expository", (0, 9, 180)),
"gutenberg_books": ("books", (0, 8, 70)),
"poetry": ("poetry", (0, 1, 10)),
"cnn": ("news", (0, 1, 20)),
}
# Complexity tiers per size: reference vocab size, max Flesch-Kincaid grade,
# document length range (words), max average words per dialogue turn.
TIERS = {
"10M": dict(vocab=8000, fk=5, fk_exp=9, doc_max=600, turn=25),
"100M": dict(vocab=15000, fk=8, fk_exp=10, doc_max=1200, turn=40),
"1000M": dict(vocab=30000, fk=12, fk_exp=12, doc_max=2000, turn=80),
}
# fk_exp: expository text (Cosmopedia textbooks) has median grade ~11 even for school
# audiences, so fk=5 passed 0.2% of it. Vocabulary coverage still guards simplicity.
# Assistant chats may average this many times the per-turn word cap of a named dialogue.
CHAT_TURN_MULT = 4
# --- run safety nets (see build.py) ---
# Relaxation ladder: (extra Flesch-Kincaid grades, change to vocabulary coverage).
LADDER = [(0, 0.0), (1, -0.02), (2, -0.04), (3, -0.06)]
PROBE_DOCS = 1000 # docs sampled before a source starts, to pick the starting ladder step
WINDOW = 5000 # docs per yield check while a source runs
MAX_SCAN = 40 # hard cap: raw words scanned per source = MAX_SCAN x its quota
# When a source falls short, same-genre substitutes make up the difference in order.
# A substitute may add at most max(EXTRA_FRAC x its quota, EXTRA_FLOOR x total target)
# and never end above SHARE_CAP x total target.
EXTRA_FRAC, EXTRA_FLOOR, SHARE_CAP = 0.5, 0.05, 0.35
FALLBACKS = {
"tinystories": ["simplestories", "cosmo_stories"],
"simplestories": ["tinystories", "cosmo_stories"],
"cosmo_stories": ["simplestories", "tinystories"],
"tinydialogues": ["soda", "smoltalk"],
"soda": ["tinydialogues", "ultrachat", "smoltalk"],
"ultrachat": ["smoltalk", "soda"],
"smoltalk": ["ultrachat", "soda"],
"tinyhelen": ["tiny_textbooks", "cosmo_v2", "fineweb_edu"],
"simplewiki": ["tiny_textbooks", "tinyhelen", "cosmo_v2"],
"enwiki_intro": ["cosmo_v2", "tiny_textbooks", "fineweb_edu"],
"cosmo_v2": ["tiny_textbooks", "tinyhelen", "fineweb_edu"],
"cosmo_v01": ["tiny_textbooks", "cosmo_v2", "tinyhelen"],
"fineweb_edu": ["cosmo_v2", "tiny_textbooks", "tinyhelen"],
"gutenberg_books": ["tinyhelen", "fineweb_edu"],
"poetry": [],
"cnn": ["fineweb_edu", "cosmo_v2"],
}
# min share of (lowercase) tokens inside the reference vocab. The original 95% @ top-5k
# passed only 3% of SimpleStories and 35% of TinyStories, so it is 90% with larger vocabs.
VOCAB_COVERAGE = 0.90
# English-only by construction and made of one-word toddler turns, which fastText misreads
LID_SKIP = {"tinydialogues"}
DOC_MIN = 80
|