Upload dataset_builder.py with huggingface_hub
Browse files- dataset_builder.py +437 -0
dataset_builder.py
ADDED
|
@@ -0,0 +1,437 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Dataset Builder v3 — Football Prediction Extractor
|
| 3 |
+
- Always outputs JSON array (even single tip)
|
| 4 |
+
- 70% single-tip / 30% multi-tip (2-4 events)
|
| 5 |
+
- Noise: random emojis, typos, missing fields, varied separators
|
| 6 |
+
- Varied date formats, bookmakers, times, headers
|
| 7 |
+
- Pure stdlib — no pip installs needed
|
| 8 |
+
"""
|
| 9 |
+
|
| 10 |
+
import csv
|
| 11 |
+
import json
|
| 12 |
+
import random
|
| 13 |
+
from pathlib import Path
|
| 14 |
+
from collections import defaultdict
|
| 15 |
+
|
| 16 |
+
# ─────────────────────────────────────────────
|
| 17 |
+
# CONFIG
|
| 18 |
+
# ─────────────────────────────────────────────
|
| 19 |
+
TEAMS_CSV = "teams_tier1_tier2.csv"
|
| 20 |
+
OUTPUT_TRAIN = "train_dataset.jsonl"
|
| 21 |
+
OUTPUT_VAL = "val_dataset.jsonl"
|
| 22 |
+
EXAMPLES_COUNT = 300
|
| 23 |
+
VAL_SPLIT = 0.1
|
| 24 |
+
|
| 25 |
+
# ─────────────────────────────────────────────
|
| 26 |
+
# SYSTEM PROMPT — always array
|
| 27 |
+
# ─────────────────────────────────────────────
|
| 28 |
+
SYSTEM_PROMPT = (
|
| 29 |
+
"You are a football data extraction assistant. "
|
| 30 |
+
"Extract structured data from the message and return ONLY a valid JSON array. "
|
| 31 |
+
"Each object in the array must have exactly these keys: "
|
| 32 |
+
"league, team_1, team_2, prediction, date, odds. "
|
| 33 |
+
"If a field is missing, use null. No extra text, no markdown."
|
| 34 |
+
)
|
| 35 |
+
|
| 36 |
+
# ─────────────────────────────────────────────
|
| 37 |
+
# VOCABULARY
|
| 38 |
+
# ─────────────────────────────────────────────
|
| 39 |
+
PREDICTIONS = [
|
| 40 |
+
"Over 1.5", "Over 2.5", "Over 3.5",
|
| 41 |
+
"Under 2.5", "Under 3.5",
|
| 42 |
+
"1X", "X2", "12",
|
| 43 |
+
"Home Win", "Away Win", "Draw",
|
| 44 |
+
"Both Teams to Score",
|
| 45 |
+
"Home Win or Draw",
|
| 46 |
+
"Away Win or Draw",
|
| 47 |
+
"GG", "NG",
|
| 48 |
+
]
|
| 49 |
+
|
| 50 |
+
DATE_FORMATS = [
|
| 51 |
+
lambda d, m, y: f"{d:02d}/{m:02d}/{y}",
|
| 52 |
+
lambda d, m, y: f"{d:02d}-{m:02d}-{y}",
|
| 53 |
+
lambda d, m, y: f"{d:02d}.{m:02d}.{y}",
|
| 54 |
+
lambda d, m, y: f"{['Jan','Feb','Mar','Apr','May','Jun','Jul','Aug','Sep','Oct','Nov','Dec'][m-1]} {d}, {y}",
|
| 55 |
+
lambda d, m, y: f"{d} {['January','February','March','April','May','June','July','August','September','October','November','December'][m-1]} {y}",
|
| 56 |
+
]
|
| 57 |
+
|
| 58 |
+
TIMES = ["13:00", "15:00", "16:00", "17:00", "18:00", "19:00", "19:45", "20:00", "20:45", "21:00", "21:45"]
|
| 59 |
+
BOOKS = ["BETANO", "Bet365", "William Hill", "Unibet", "1xBet", "Betway", "Bwin", "Pinnacle"]
|
| 60 |
+
HEADERS = ["Prediction of the Day", "Football Tip", "Best Bet Today", "Daily Pick", "Top Prediction", "Sure Tip", "VIP Prediction"]
|
| 61 |
+
SEPARATORS = [" - ", " vs ", " v ", " – ", " VS ", " x "]
|
| 62 |
+
EXTRA_EMOJIS = ["🔥","💥","🎯","👀","💰","🏅","⚡️","🙌","👇","✨","📊","💎","🤑","🚨","✅","❇️","🆕","📌","👑","🃏"]
|
| 63 |
+
|
| 64 |
+
MULTI_HEADERS = [
|
| 65 |
+
"⚽️ 𝐏𝐫𝐞𝐝𝐢𝐜𝐭𝐢𝐨𝐧𝐬 𝐨𝐟 𝐭𝐡𝐞 𝐃𝐚𝐲 ⚽️",
|
| 66 |
+
"🔥 TODAY'S FOOTBALL TIPS 🔥",
|
| 67 |
+
"💰 Daily Predictions 💰",
|
| 68 |
+
"⚡️ Best Bets Today ⚡️",
|
| 69 |
+
"📊 Football Tips",
|
| 70 |
+
"🎯 Today's Picks",
|
| 71 |
+
]
|
| 72 |
+
|
| 73 |
+
MULTI_FOOTERS = [
|
| 74 |
+
"For more predictions visit www.eaglepredict.com",
|
| 75 |
+
"Follow us for daily tips! 🙌",
|
| 76 |
+
"Good luck everyone! 🍀",
|
| 77 |
+
"Join our VIP channel for more! 💎",
|
| 78 |
+
"Win big today! 🤑",
|
| 79 |
+
"", # no footer sometimes
|
| 80 |
+
]
|
| 81 |
+
|
| 82 |
+
# ─────────────────────────────────────────────
|
| 83 |
+
# SINGLE TIP TEMPLATES
|
| 84 |
+
# placeholders: {league} {team_1} {team_2} {prediction}
|
| 85 |
+
# {date} {odds} {time} {header} {book} {sep}
|
| 86 |
+
# templates 7 and 8 intentionally omit odds/date
|
| 87 |
+
# ─────────────────────────────────────────────
|
| 88 |
+
SINGLE_TEMPLATES = [
|
| 89 |
+
# 1 structured Telegram bold style
|
| 90 |
+
"⚽️ {header} ⚽️\n𝐃𝐚𝐭𝐞: {date}\n𝐋𝐞𝐚𝐠𝐮𝐞: {league}\n𝐌𝐚𝐭𝐜𝐡: {team_1}{sep}{team_2}\n𝐊𝐢𝐜𝐤 𝐨𝐟𝐟: {time} WAT\n✅{prediction}\n✅Odds @{odds} on {book}",
|
| 91 |
+
|
| 92 |
+
# 2 plain structured
|
| 93 |
+
"⚽️ {header} ⚽️\nDate: {date}\nLeague: {league}\nMatch: {team_1}{sep}{team_2}\nKick off: {time} WAT\n✅{prediction}\n✅Odds @{odds} on {book}",
|
| 94 |
+
|
| 95 |
+
# 3 emoji compact
|
| 96 |
+
"🏆 {league}\n{team_1}{sep}{team_2}\n📅 {date} | ⏰ {time}\n🎯 Tip: {prediction}\n💰 Odds: {odds}",
|
| 97 |
+
|
| 98 |
+
# 4 casual noisy
|
| 99 |
+
"wow predictions present\nINCREDIBLE MATCH BETWEEN {team_1}{sep}{team_2}\nTime: {time}\nwe forecast {prediction}\nOdds {odds}",
|
| 100 |
+
|
| 101 |
+
# 5 one-liner
|
| 102 |
+
"{team_1}{sep}{team_2} | {league} | {date} | {prediction} @ {odds}",
|
| 103 |
+
|
| 104 |
+
# 6 verbose channel
|
| 105 |
+
"🔥 Today's football tip 🔥\nCompetition: {league}\nGame: {team_1}{sep}{team_2}\nDate: {date}, KO {time}\nOur pick: {prediction}\nBest odds: {odds} ({book})\nGood luck! ⚽",
|
| 106 |
+
|
| 107 |
+
# 7 minimal no emojis
|
| 108 |
+
"Match: {team_1}{sep}{team_2}\nLeague: {league}\nDate: {date}\nPrediction: {prediction}\nOdds: {odds}",
|
| 109 |
+
|
| 110 |
+
# 8 different field order
|
| 111 |
+
"📆 {date} | {time}\n⚽ {league}: {team_1}{sep}{team_2}\n✔️ {prediction} | @{odds}",
|
| 112 |
+
|
| 113 |
+
# 9 ALL CAPS noisy
|
| 114 |
+
"MATCH: {team_1}{sep}{team_2}\nLEAGUE: {league}\nDATE: {date}\nPICK: {prediction}\nODDS: {odds}",
|
| 115 |
+
|
| 116 |
+
# 10 missing odds intentionally
|
| 117 |
+
"⚽️ {header}\n{league}\n{team_1}{sep}{team_2}\n{date}\nPrediction: {prediction}",
|
| 118 |
+
|
| 119 |
+
# 11 missing date intentionally
|
| 120 |
+
"🏟️ {league}\n{team_1}{sep}{team_2}\nTip: {prediction}\nOdds: {odds} on {book}",
|
| 121 |
+
|
| 122 |
+
# 12 missing league intentionally
|
| 123 |
+
"{team_1}{sep}{team_2}\n📅 {date}\n✅ {prediction} @ {odds}",
|
| 124 |
+
|
| 125 |
+
# 13 telegram minimal
|
| 126 |
+
"📌 {league}\n{team_1}{sep}{team_2} — {date}\n{prediction} | {odds}",
|
| 127 |
+
|
| 128 |
+
# 14 with extra commentary noise
|
| 129 |
+
"Today I really like this match 👇\n{team_1}{sep}{team_2} ({league})\nDate: {date}\nMy pick: {prediction}\nOdds: {odds} on {book}",
|
| 130 |
+
]
|
| 131 |
+
|
| 132 |
+
# ─────────────────────────────────────────────
|
| 133 |
+
# MULTI-TIP BLOCK TEMPLATES (per tip)
|
| 134 |
+
# extra placeholder: {n} = tip number
|
| 135 |
+
# ─────────────────────────────────────────────
|
| 136 |
+
MULTI_BLOCK_TEMPLATES = [
|
| 137 |
+
# Telegram numbered bold
|
| 138 |
+
"⚽️ 𝗙𝗼𝗼𝘁𝗯𝗮𝗹𝗹 𝗧𝗶𝗽 {n} ⚽️\n𝐃𝐚𝐭𝐞: {date}\n𝐋𝐞𝐚𝐠𝐮𝐞: {league}\n𝐌𝐚𝐭𝐜𝐡: {team_1}{sep}{team_2}\n𝐊𝐢𝐜𝐤 𝐨𝐟𝐟: {time} WAT\n✅{prediction}\n✅Odds @{odds} on {book}",
|
| 139 |
+
|
| 140 |
+
# plain numbered
|
| 141 |
+
"Tip {n}:\nLeague: {league}\nMatch: {team_1}{sep}{team_2}\nDate: {date} | KO: {time}\nPrediction: {prediction} @ {odds}",
|
| 142 |
+
|
| 143 |
+
# compact numbered
|
| 144 |
+
"#{n} {league} | {team_1}{sep}{team_2} | {date}\n→ {prediction} @ {odds}",
|
| 145 |
+
|
| 146 |
+
# emoji numbered
|
| 147 |
+
"🎯 Pick #{n}\n{team_1}{sep}{team_2} ({league})\n📅 {date} ⏰ {time}\n✅ {prediction} | odds: {odds}",
|
| 148 |
+
|
| 149 |
+
# minimal numbered
|
| 150 |
+
"{n}. {team_1}{sep}{team_2} — {league} — {prediction} @ {odds} ({date})",
|
| 151 |
+
]
|
| 152 |
+
|
| 153 |
+
# ─────────────────────────────────────────────
|
| 154 |
+
# LOAD TEAMS FROM CSV
|
| 155 |
+
# ─────────────────────────────────────────────
|
| 156 |
+
def load_teams(csv_path: str) -> dict:
|
| 157 |
+
leagues = defaultdict(list)
|
| 158 |
+
path = Path(csv_path)
|
| 159 |
+
if not path.exists():
|
| 160 |
+
raise FileNotFoundError(f"CSV not found: {csv_path}")
|
| 161 |
+
with open(path, encoding="utf-8") as f:
|
| 162 |
+
sample = f.read(2048)
|
| 163 |
+
f.seek(0)
|
| 164 |
+
delimiter = "\t" if "\t" in sample else ","
|
| 165 |
+
reader = csv.DictReader(f, delimiter=delimiter)
|
| 166 |
+
for row in reader:
|
| 167 |
+
row = {k.strip(): v.strip() for k, v in row.items()}
|
| 168 |
+
country = row.get("Country", "")
|
| 169 |
+
league = row.get("League", "")
|
| 170 |
+
team = row.get("Team", "")
|
| 171 |
+
if country and league and team:
|
| 172 |
+
leagues[(country, league)].append(team)
|
| 173 |
+
total = sum(len(v) for v in leagues.values())
|
| 174 |
+
print(f"[✓] Loaded {total} teams across {len(leagues)} leagues")
|
| 175 |
+
return leagues
|
| 176 |
+
|
| 177 |
+
# ─────────────────────────────────────────────
|
| 178 |
+
# RANDOM HELPERS
|
| 179 |
+
# ─────────────────────────────────────────────
|
| 180 |
+
def random_date() -> str:
|
| 181 |
+
month = random.randint(8, 12) if random.random() < 0.5 else random.randint(1, 5)
|
| 182 |
+
year = 2025 if month >= 8 else 2026
|
| 183 |
+
day = random.randint(1, 28)
|
| 184 |
+
return random.choice(DATE_FORMATS)(day, month, year)
|
| 185 |
+
|
| 186 |
+
def random_odds() -> float:
|
| 187 |
+
return round(random.uniform(1.05, 3.50), 2)
|
| 188 |
+
|
| 189 |
+
def random_fixture(leagues: dict) -> dict | None:
|
| 190 |
+
key = random.choice(list(leagues.keys()))
|
| 191 |
+
teams = leagues[key]
|
| 192 |
+
if len(teams) < 2:
|
| 193 |
+
return None
|
| 194 |
+
_, league = key
|
| 195 |
+
team_1, team_2 = random.sample(teams, 2)
|
| 196 |
+
return {
|
| 197 |
+
"league": league,
|
| 198 |
+
"team_1": team_1,
|
| 199 |
+
"team_2": team_2,
|
| 200 |
+
"prediction": random.choice(PREDICTIONS),
|
| 201 |
+
"date": random_date(),
|
| 202 |
+
"odds": random_odds(),
|
| 203 |
+
}
|
| 204 |
+
|
| 205 |
+
# ─────────────────────────────────────────────
|
| 206 |
+
# NOISE FUNCTIONS
|
| 207 |
+
# ─────────────────────────────────────────────
|
| 208 |
+
def inject_emojis(text: str) -> str:
|
| 209 |
+
"""40% chance: sprinkle 1-3 random emojis into random lines."""
|
| 210 |
+
if random.random() < 0.40:
|
| 211 |
+
emojis = random.sample(EXTRA_EMOJIS, k=random.randint(1, 3))
|
| 212 |
+
lines = text.split("\n")
|
| 213 |
+
for e in emojis:
|
| 214 |
+
idx = random.randint(0, len(lines) - 1)
|
| 215 |
+
lines[idx] = (e + " " + lines[idx]) if random.random() < 0.5 else (lines[idx] + " " + e)
|
| 216 |
+
return "\n".join(lines)
|
| 217 |
+
return text
|
| 218 |
+
|
| 219 |
+
def inject_typos(text: str) -> str:
|
| 220 |
+
"""15% chance: swap two adjacent chars in a random word."""
|
| 221 |
+
if random.random() < 0.15:
|
| 222 |
+
words = text.split(" ")
|
| 223 |
+
idx = random.randint(0, len(words) - 1)
|
| 224 |
+
w = words[idx]
|
| 225 |
+
if len(w) > 3 and w.isalpha():
|
| 226 |
+
i = random.randint(0, len(w) - 2)
|
| 227 |
+
w = w[:i] + w[i+1] + w[i] + w[i+2:]
|
| 228 |
+
words[idx] = w
|
| 229 |
+
return " ".join(words)
|
| 230 |
+
return text
|
| 231 |
+
|
| 232 |
+
def inject_extra_lines(text: str) -> str:
|
| 233 |
+
"""20% chance: add irrelevant noise lines."""
|
| 234 |
+
noise_lines = [
|
| 235 |
+
"For more predictions visit www.eaglepredict.com",
|
| 236 |
+
"Join our VIP channel 💎",
|
| 237 |
+
"Yesterday result: WIN ✅",
|
| 238 |
+
"Record this week: 8W 2L",
|
| 239 |
+
"All tips are for 18+ only",
|
| 240 |
+
"Use responsible gambling 🙏",
|
| 241 |
+
]
|
| 242 |
+
if random.random() < 0.20:
|
| 243 |
+
line = random.choice(noise_lines)
|
| 244 |
+
if random.random() < 0.5:
|
| 245 |
+
return line + "\n" + text
|
| 246 |
+
else:
|
| 247 |
+
return text + "\n" + line
|
| 248 |
+
return text
|
| 249 |
+
|
| 250 |
+
def maybe_null_field(fixture: dict, has_odds: bool, has_date: bool, has_league: bool) -> dict:
|
| 251 |
+
"""
|
| 252 |
+
Randomly null out one field (20% chance).
|
| 253 |
+
Respects whether template already omits it.
|
| 254 |
+
"""
|
| 255 |
+
f = dict(fixture)
|
| 256 |
+
if not has_odds:
|
| 257 |
+
f["odds"] = None
|
| 258 |
+
if not has_date:
|
| 259 |
+
f["date"] = None
|
| 260 |
+
if not has_league:
|
| 261 |
+
f["league"] = None
|
| 262 |
+
# extra random null on top
|
| 263 |
+
if random.random() < 0.20:
|
| 264 |
+
field = random.choice(["odds", "date", "league"])
|
| 265 |
+
f[field] = None
|
| 266 |
+
return f
|
| 267 |
+
|
| 268 |
+
def apply_noise(text: str) -> str:
|
| 269 |
+
text = inject_emojis(text)
|
| 270 |
+
text = inject_typos(text)
|
| 271 |
+
text = inject_extra_lines(text)
|
| 272 |
+
return text
|
| 273 |
+
|
| 274 |
+
# ─────────────────────────────────────────────
|
| 275 |
+
# EXAMPLE GENERATORS
|
| 276 |
+
# ─────────────────────────────────────────────
|
| 277 |
+
def make_single_example(leagues: dict) -> dict | None:
|
| 278 |
+
fixture = random_fixture(leagues)
|
| 279 |
+
if not fixture:
|
| 280 |
+
return None
|
| 281 |
+
|
| 282 |
+
template = random.choice(SINGLE_TEMPLATES)
|
| 283 |
+
has_odds = "{odds}" in template
|
| 284 |
+
has_date = "{date}" in template
|
| 285 |
+
has_league= "{league}" in template
|
| 286 |
+
sep = random.choice(SEPARATORS)
|
| 287 |
+
|
| 288 |
+
input_text = template.format(
|
| 289 |
+
sep = sep,
|
| 290 |
+
league = fixture["league"],
|
| 291 |
+
team_1 = fixture["team_1"],
|
| 292 |
+
team_2 = fixture["team_2"],
|
| 293 |
+
prediction = fixture["prediction"],
|
| 294 |
+
date = fixture["date"],
|
| 295 |
+
odds = fixture["odds"],
|
| 296 |
+
time = random.choice(TIMES),
|
| 297 |
+
header = random.choice(HEADERS),
|
| 298 |
+
book = random.choice(BOOKS),
|
| 299 |
+
)
|
| 300 |
+
input_text = apply_noise(input_text)
|
| 301 |
+
output_json = maybe_null_field(fixture, has_odds, has_date, has_league)
|
| 302 |
+
|
| 303 |
+
return {
|
| 304 |
+
"input": input_text,
|
| 305 |
+
"output": [output_json], # always array
|
| 306 |
+
}
|
| 307 |
+
|
| 308 |
+
|
| 309 |
+
def make_multi_example(leagues: dict) -> dict | None:
|
| 310 |
+
n_tips = random.randint(2, 4)
|
| 311 |
+
fixtures = [random_fixture(leagues) for _ in range(n_tips * 2)]
|
| 312 |
+
fixtures = [f for f in fixtures if f][:n_tips]
|
| 313 |
+
if len(fixtures) < 2:
|
| 314 |
+
return None
|
| 315 |
+
|
| 316 |
+
block_template = random.choice(MULTI_BLOCK_TEMPLATES)
|
| 317 |
+
sep = random.choice(SEPARATORS)
|
| 318 |
+
blocks = []
|
| 319 |
+
|
| 320 |
+
for i, f in enumerate(fixtures, 1):
|
| 321 |
+
has_odds = "{odds}" in block_template
|
| 322 |
+
has_date = "{date}" in block_template
|
| 323 |
+
has_league= "{league}" in block_template
|
| 324 |
+
block = block_template.format(
|
| 325 |
+
n = i,
|
| 326 |
+
sep = sep,
|
| 327 |
+
league = f["league"],
|
| 328 |
+
team_1 = f["team_1"],
|
| 329 |
+
team_2 = f["team_2"],
|
| 330 |
+
prediction = f["prediction"],
|
| 331 |
+
date = f["date"],
|
| 332 |
+
odds = f["odds"],
|
| 333 |
+
time = random.choice(TIMES),
|
| 334 |
+
book = random.choice(BOOKS),
|
| 335 |
+
)
|
| 336 |
+
blocks.append((block, f, has_odds, has_date, has_league))
|
| 337 |
+
|
| 338 |
+
header = random.choice(MULTI_HEADERS)
|
| 339 |
+
footer = random.choice(MULTI_FOOTERS)
|
| 340 |
+
parts = [header] + [b[0] for b in blocks] + ([footer] if footer else [])
|
| 341 |
+
input_text = "\n".join(parts)
|
| 342 |
+
input_text = apply_noise(input_text)
|
| 343 |
+
|
| 344 |
+
output = [
|
| 345 |
+
maybe_null_field(f, has_odds, has_date, has_league)
|
| 346 |
+
for _, f, has_odds, has_date, has_league in blocks
|
| 347 |
+
]
|
| 348 |
+
|
| 349 |
+
return {"input": input_text, "output": output}
|
| 350 |
+
|
| 351 |
+
# ─────────────────────────────────────────────
|
| 352 |
+
# FORMAT AS TRAINING EXAMPLE
|
| 353 |
+
# ──────────────────��──────────────────────────
|
| 354 |
+
def make_training_example(ex: dict) -> dict:
|
| 355 |
+
return {
|
| 356 |
+
"messages": [
|
| 357 |
+
{"role": "system", "content": SYSTEM_PROMPT},
|
| 358 |
+
{"role": "user", "content": ex["input"].strip()},
|
| 359 |
+
{"role": "assistant", "content": json.dumps(ex["output"], ensure_ascii=False)},
|
| 360 |
+
]
|
| 361 |
+
}
|
| 362 |
+
|
| 363 |
+
# ─────────────────────────────────────────────
|
| 364 |
+
# MAIN
|
| 365 |
+
# ─────────────────────────────────────────────
|
| 366 |
+
def build_dataset():
|
| 367 |
+
leagues = load_teams(TEAMS_CSV)
|
| 368 |
+
examples = []
|
| 369 |
+
|
| 370 |
+
n_single = int(EXAMPLES_COUNT * 0.70)
|
| 371 |
+
n_multi = EXAMPLES_COUNT - n_single
|
| 372 |
+
print(f"[1/2] Generating {n_single} single-tip + {n_multi} multi-tip examples...")
|
| 373 |
+
|
| 374 |
+
# single tip
|
| 375 |
+
attempts = 0
|
| 376 |
+
while len([e for e in examples if len(json.loads(e["messages"][2]["content"])) == 1]) < n_single:
|
| 377 |
+
attempts += 1
|
| 378 |
+
if attempts > n_single * 5:
|
| 379 |
+
break
|
| 380 |
+
ex = make_single_example(leagues)
|
| 381 |
+
if ex:
|
| 382 |
+
examples.append(make_training_example(ex))
|
| 383 |
+
|
| 384 |
+
# multi tip
|
| 385 |
+
attempts = 0
|
| 386 |
+
while len([e for e in examples if len(json.loads(e["messages"][2]["content"])) > 1]) < n_multi:
|
| 387 |
+
attempts += 1
|
| 388 |
+
if attempts > n_multi * 5:
|
| 389 |
+
break
|
| 390 |
+
ex = make_multi_example(leagues)
|
| 391 |
+
if ex:
|
| 392 |
+
examples.append(make_training_example(ex))
|
| 393 |
+
|
| 394 |
+
print(f" → {len(examples)} total examples generated")
|
| 395 |
+
|
| 396 |
+
# ── Write files ────────────────────────────
|
| 397 |
+
print("[2/2] Writing dataset files...")
|
| 398 |
+
random.shuffle(examples)
|
| 399 |
+
split = int(len(examples) * (1 - VAL_SPLIT))
|
| 400 |
+
train, val = examples[:split], examples[split:]
|
| 401 |
+
|
| 402 |
+
for path, data in [(OUTPUT_TRAIN, train), (OUTPUT_VAL, val)]:
|
| 403 |
+
with open(path, "w", encoding="utf-8") as f:
|
| 404 |
+
for ex in data:
|
| 405 |
+
f.write(json.dumps(ex, ensure_ascii=False) + "\n")
|
| 406 |
+
|
| 407 |
+
# ── Stats ──────────────────────────────────
|
| 408 |
+
all_ex = train + val
|
| 409 |
+
single = sum(1 for e in all_ex if len(json.loads(e["messages"][2]["content"])) == 1)
|
| 410 |
+
multi = len(all_ex) - single
|
| 411 |
+
nulls = sum(
|
| 412 |
+
1 for e in all_ex
|
| 413 |
+
for obj in json.loads(e["messages"][2]["content"])
|
| 414 |
+
if any(v is None for v in obj.values())
|
| 415 |
+
)
|
| 416 |
+
|
| 417 |
+
print(f"\n✅ Done!")
|
| 418 |
+
print(f" {OUTPUT_TRAIN} → {len(train)} examples")
|
| 419 |
+
print(f" {OUTPUT_VAL} → {len(val)} examples")
|
| 420 |
+
print(f" Single-tip → {single}")
|
| 421 |
+
print(f" Multi-tip → {multi}")
|
| 422 |
+
print(f" With null fields→ {nulls}")
|
| 423 |
+
|
| 424 |
+
# ── Previews ───────────────────────────────
|
| 425 |
+
print("\n── Single-tip sample ───────────────────────")
|
| 426 |
+
s = next(e for e in examples if len(json.loads(e["messages"][2]["content"])) == 1)
|
| 427 |
+
for msg in s["messages"]:
|
| 428 |
+
print(f"[{msg['role']}]\n{msg['content'][:200]}\n")
|
| 429 |
+
|
| 430 |
+
print("── Multi-tip sample ────────────────────────")
|
| 431 |
+
m = next(e for e in examples if len(json.loads(e["messages"][2]["content"])) > 1)
|
| 432 |
+
for msg in m["messages"]:
|
| 433 |
+
print(f"[{msg['role']}]\n{msg['content'][:300]}\n")
|
| 434 |
+
|
| 435 |
+
|
| 436 |
+
if __name__ == "__main__":
|
| 437 |
+
build_dataset()
|