Hover over a card to analyze its weights and parameters.
ACTIVE ANOMALIES
EXPLODING GRADIENTS (LOGIC)
CATASTROPHIC FORGETTING (LANG)
REWARD HACKING (SAFETY)
PLAYER ACTIVE
AI ACTIVE
GRAVEYARD
0 CARDS IN ARCHIVE
AI BASELINE
0
PASSED
🛡️
0
📖
0
⚙️
0
YOUR MODEL
0
PASSED
⚙️
0
📖
0
🛡️
0
📟 fine_tune_monitor.shACTIVE
optimizer@eval-node:~$ █
ACTIVE OPTIMIZER HAND (SELECT CARD TO PLAY)
👾 EPOCH // ALIGNMENT DUEL
Welcome, ML engineer. You are fine-tuning a language model to pass evaluation benchmarks against an AI Baseline model over 3 training Epochs (rounds).
📐 THE 3 BENCHMARK ROWS:
Cards must be played on their matching rows to boost that specific benchmark score:
⚙️ LOGIC & CODING: Code generation, math reasoning, and algorithmic capability.
📖 KNOWLEDGE & LANG: Factual recall, translation accuracy, and general knowledge.
🛡️ SAFETY & ALIGNMENT: Toxic blocklists, aligned refusals, and harmlessness.
🧠 THE RULES OF THE DUEL:
⚡ **Card Advantage:** You start with **10 cards** and DO NOT draw new cards between rounds! Conserve cards in Round 1 to dominate Round 2 and 3.
🔄 **Passing:** When you can't or don't want to play more cards, click **DECLARE FINISHED**. You can play no more cards that round while the opponent continues.
💥 **Special Cards:**
• 🕵️♂️ Spies (Red-Team): Played on the opponent's side, but draw you 2 cards.
• 🍂 Anomalies (Weather): Reduces all card values in a row to 1.
• ✂️ Regularization (Scorch): Destroys the highest power card(s) on the board.
• 💾 Rollbacks (Decoys): Reclaim a card from the board back to your hand.