t1-strategy-arena-frozenlake-baseline-together_ai-moonshotai-kimi-k2-instruct-78cc6a2e

Strategy compliance baseline — FrozenLake arena evaluation.

No strategy instruction was given. This is the baseline to observe the model's natural reasoning patterns across game turns. Per-turn reasoning is extracted from transcripts for manual review.

Results

Metric	Value
Win rate	10/10 (100.0%)
Avg turns	3.5
Errors	0

Details

Parameter	Value
Model	together_ai/moonshotai/Kimi-K2-Instruct-0905
Thinking	False
Temperature	0.7
Max tokens	4096
Games	FrozenLake-v0, FrozenLake-v0, FrozenLake-v0, FrozenLake-v0, FrozenLake-v0, FrozenLake-v0, FrozenLake-v0, FrozenLake-v0, FrozenLake-v0, FrozenLake-v0
Procedural boards	True
Seed	42
Variant	baseline

Columns

• transcript: Full game transcript (JSON array of turn objects)
• per_turn_reasoning: Extracted reasoning per turn (JSON array of strings)
• reward: Final game reward (1.0 = win, negative = loss)
• num_turns: Number of turns played

Usage

from datasets import load_dataset
import json

ds = load_dataset("reasoning-degeneration-dev/t1-strategy-arena-frozenlake-baseline-together_ai-moonshotai-kimi-k2-instruct-78cc6a2e", split="train")
for row in ds:
    reasoning = json.loads(row["per_turn_reasoning"])
    print(f"Game {row['game_id']}: {len(reasoning)} turns, reward={row['reward']}")
    for t, r in enumerate(reasoning):
        if r.strip():
            print(f"  Turn {t}: {r[:200]}...")

---

Tracked in reasoning-degeneration-dev/PROJECT-MANIFEST