SlimeBench-balanced

Action-balanced behavioral cloning of the deterministic SlimeVolley CMA-ES teacher. There are 64 rows, each with 2048 training and 100 test conversations. Each conversation comes from its own independently seeded complete rollout. Seeds are globally disjoint across rows and splits and distinct from the original SlimeBench production run. No examples are duplicated to fill action quotas.

Each row and split has nearly equal counts of the six underlying actions: 341 or 342 training examples per action, and 16 or 17 test examples per action. Remainder recipients and example order are randomized independently per split. For each assigned action, an independent teacher rollout is completed and reservoir sampling selects uniformly among timesteps where the teacher chooses that action. If the action never occurs, the rollout is discarded and a fresh seed is tried. Rollouts are not selected by reward, wins, or episode length.

This samples eligible rollouts equally, then matching timesteps uniformly; it does not pool all matching timesteps across games. It intentionally changes the natural state/action distribution. Evaluation measures action-balanced agreement with this teacher, not closed-loop reward or optimality.

The six effective actions are no movement, forward, forward+jump, jump, backward+jump, and backward. For each row, their A–F labels and the order of the 12 observation coordinates are independently randomly permuted. Both mappings are fixed within the row and unique across rows. Conversations contain only an anonymous numeric list and a single-letter answer, with no system instruction. Every published action is checked against the teacher on the serialized float32 input. Original feature names, private permutations, seeds, and rollout metadata are kept locally and excluded from the published dataset.

Public columns: row_id, num_train, num_test, train_data, test_data. The Hub split train contains the 64 outer rows; each row contains its own training and held-out lists of two-message conversations.

Reproduce

Use Python 3.11, NumPy 1.26.4, Gym 0.23.1, datasets 4.4.1, pyarrow 21.0.0, huggingface-hub 0.36.0, python-dotenv, and opencv-python-headless 4.10.0.84. Clone SlimeVolleyGym at revision 8ac22434fc5a587a3395311f7854aeda17d10303. The original feedforward weights are zoo/cmaes/slimevolley.cma.64.96.best.json, SHA-256 3523cca6459ea6a979b98b19de8466eba8d5d1318ce6fa71f8f7cd084f38ce9b. The opponent is the built-in recurrent BaselinePolicy, reset every rollout. Positive teacher outputs press buttons; simultaneous opposite directions cancel. SlimeVolleyGym is distributed under Apache-2.0.

Both slime_bench_balanced.py and its helper slime_bench.py are provided with the dataset. In the outer-loop repository:

OPENBLAS_NUM_THREADS=1 /tmp/slimebench-venv/bin/python \
  data_preparation/benchmarks/slime_bench_balanced.py \
  --source-dir /tmp/slimebench-upstream \
  --output-dir local_data/slime_bench_balanced --workers 32 --upload \
  --output-dataset aklein4/SlimeBench-balanced

The default seed is 10000000. Completed rows are checkpointed for resumption; the manifest guards against changed source/settings. Publication requires the full production size. The validation report includes split action counts, unique rollout count, rejected rollout count, and dataset checksum.

from datasets import load_dataset
rows = load_dataset("aklein4/SlimeBench-balanced", split="train")

Validation

{
  "rows": 64,
  "conversations": 137472,
  "distinct_rollouts": 137472,
  "attempted_rollouts": 137472,
  "rejected_rollouts": 0,
  "cross_split_duplicate_prompts": 0,
  "mean_rollout_return": 1.1095204841713222,
  "action_counts": {
    "train": [
      21850,
      21839,
      21843,
      21839,
      21852,
      21849
    ],
    "test": [
      1066,
      1064,
      1071,
      1068,
      1072,
      1059
    ]
  },
  "jsonl_sha256": "39a7b12517a587e6ed00d291a7a84e362f506cbfe29dadfdb6d2330bb7eb035e"
}