README.md

---
dataset_info:
  features:
  - name: row_id
    dtype: int64
  - name: num_train
    dtype: int64
  - name: num_test
    dtype: int64
  - name: train_data
    list:
      list:
      - name: role
        dtype: string
      - name: content
        dtype: string
  - name: test_data
    list:
      list:
      - name: role
        dtype: string
      - name: content
        dtype: string
  splits:
  - name: train
    num_bytes: 22338945
    num_examples: 64
  download_size: 8940626
  dataset_size: 22338945
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# SlimeBench

SlimeBench measures offline imitation of a deterministic SlimeVolley controller.
It contains 64 rows, each with 2,048 training and 100 test conversations. Each
conversation comes from a different independently seeded, complete rollout:
137,472 rollouts total, with disjoint seeds across every row and split.

A conversation contains only a user message with a bracketed list of 12 numbers
and an assistant message with one letter from `A` through `F`. There are no feature
names, action descriptions, or system messages. Each row has its own random
permutation of observation coordinates and assignment of the six effective actions
to letters. Both mappings are fixed within the row and distinct across all rows.
The public columns are `row_id`, `num_train`, `num_test`, `train_data`, `test_data`;
the latter two contain lists of two-message conversations.

## Teacher and sampling

The teacher is the upstream feedforward CMA-ES policy, using the original
`zoo/cmaes/slimevolley.cma.64.96.best.json` weights. It plays against the built-in
recurrent `BaselinePolicy`. Both simulator and opponent state are reset for each
rollout; upstream `env.reset()` alone does not reset the opponent's recurrent state.
The pinned upstream revision is `8ac22434fc5a587a3395311f7854aeda17d10303`:
[environment](https://github.com/hardmaru/slimevolleygym/blob/8ac22434fc5a587a3395311f7854aeda17d10303/slimevolleygym/slimevolley.py),
[feedforward policy](https://github.com/hardmaru/slimevolleygym/blob/8ac22434fc5a587a3395311f7854aeda17d10303/slimevolleygym/mlp.py).
SlimeVolleyGym is distributed under Apache-2.0.

The policy receives float32 observations, serialized with nine significant digits
so every input round-trips exactly. Positive network outputs press the three
buttons; simultaneous forward/backward presses cancel, leaving six effective
actions. These are mapped to letters without changing the rollout dynamics.
Every published label is recomputed from its serialized input before upload.

Reservoir sampling selects one timestep uniformly within each complete rollout,
including games that reach the 3,000-step limit. No action balancing, win filtering,
or loss filtering is performed. Thus the dataset preserves the teacher's natural
action frequencies under this episode-uniform sampling scheme; it does not weight
longer episodes more heavily. Sampled conversations are shuffled within each split.

Accuracy measures agreement with this teacher, not optimality or closed-loop
playing strength. Natural action frequencies can be highly imbalanced: report
majority-class baselines when interpreting accuracy. Hidden random action mappings
give expected zero-example accuracy of 1/6 over mappings, assuming the evaluator
receives no mapping information. They do not make every finite row balanced.

## Reproduce

The [generator](./slime_bench.py) uses an isolated legacy
Gym environment; it does not modify the training environment. From the outer-loop
repository root:

```bash
uv venv --python 3.11 /tmp/slimebench-venv
uv pip install --python /tmp/slimebench-venv/bin/python \
  numpy==1.26.4 gym==0.23.1 opencv-python-headless==4.10.0.84 \
  datasets==4.4.1 pyarrow==21.0.0 huggingface-hub==0.36.0 python-dotenv==1.2.4
git clone https://github.com/hardmaru/slimevolleygym.git /tmp/slimebench-upstream
git -C /tmp/slimebench-upstream checkout 8ac22434fc5a587a3395311f7854aeda17d10303
OPENBLAS_NUM_THREADS=1 /tmp/slimebench-venv/bin/python \
  data_preparation/benchmarks/slime_bench.py \
  --source-dir /tmp/slimebench-upstream \
  --output-dir local_data/slime_bench --workers 32 --upload
```

The default upload destination is `aklein4/SlimeBench`, configuration `default`,
split `train`. The 64 outer rows are uploaded together; the support/test distinction
is within each row. Authentication uses the existing Hugging Face login or `HF_TOKEN`.
Use `--rows 2 --num-train 16 --num-test 8 --workers 2` and a different output directory
without `--upload` for a smoke test.

Completed rows are checkpointed and reused on reruns with the same source and
generation settings. Worker count can change on resume. Local checkpoints contain
the private permutations, original observations, rollout seeds, sampled timesteps,
and episode returns; these must not be exposed as model inputs and are not uploaded.
The output directory also contains `slime_bench.jsonl`, `manifest.json`,
`validation.json`, and (after publication) `upload.json`.

Load conversations with:

```python
from datasets import load_dataset
rows = load_dataset("aklein4/SlimeBench", split="train")
support = rows[0]["train_data"]
test = rows[0]["test_data"]
```

## Published run

- 137,472 conversations from distinct rollouts.
- Mean teacher rollout return: 1.1126.
- Exact train/test prompt overlaps within rows: 0.
- Generation seed: 42.
- JSONL SHA-256: `b04fabf11357dc0c63b307e6799a0d4d58b35c08539c06718cbd4aa675b35242`.