fracapuano/reacher_context_split

Reacher context split

2,048 original observations sampled from fracapuano/lewm-reacher, with 1,024 observations in each context and native embeddings from the official quentinll/lewm-reacher checkpoint.

Contexts and angles

angle	Context	Wrapped q1 interval	Center	Half-width	Rows
narrow	A	[-15°, 15°)	0°	15°	1,024
broad	B	[75°, 105°)	90°	15°	1,024

narrow and broad are context names. Both intervals are 30° wide; broad is not the complement of the narrow interval. angle_split=15 records the half-width in degrees. Interval lower bounds are inclusive and upper bounds exclusive.

q1 is the first joint angle, converted from the source radians and wrapped into [-180°,180°). q2 is the relative second joint angle, converted to degrees with no wrapping or clipping. Labels are the selected observations' recorded angles.

Columns

Column	Type	Meaning
image	Image	Original RGB 224×224 image bytes
q1	float32	Wrapped first joint angle, degrees
q2	float32	Relative second joint angle, degrees
angle	string	narrow or broad
angle_split	int32	Context half-width in degrees, always 15
embedding	float32[192]	Native projected observation embedding

Sampling

Uniform random sampling of 1,024 source rows without replacement within each context, using NumPy default_rng(42). Candidate rows retain source Lance row order. Context A is sampled first and B second using the same RNG. q2 was not matched, weighted, filtered, or sampled uniformly in angle space. Images are original source frames, with no rerendering or recompression. Rows are stored in random draw order, with A followed by B.

The repository has one Hugging Face train split containing all 2,048 rows. This is a storage split, not a predefined probe training set. Adjacent source frames may be correlated; form episode-disjoint training and evaluation partitions for subsequent probes. Source episode and step indices were retained in the local sampling manifest, not added to this six-column table.

Embeddings

• Checkpoint revision: 62adae4b71dc474ddf8f794c476ebfe737a743ca
• Source dataset revision: 852681f88696cd00b0c8cd6eb3baf5eddfba1a65
• Weights SHA-256: eb70b1fd5409f8f81875d62f5ee5a20dd220a3128a477de66b5760f475f0f469
• RGB uint8 → float32 / 255 → ImageNet normalization → bilinear antialiased resize to 224.
• Single-frame model.encode({"pixels": pixels})["emb"][:, 0]: CLS token after the checkpoint's projector, in the native world-model planning space. No additional latent normalization, probe, or erasure.
• MPS, float32, model.eval(), torch.inference_mode(), batch size 64; no CPU fallback.

All labels and image bytes were checked against the selected source rows. All embeddings are finite. Local Parquet readback preserved the original image bytes, labels, and embeddings. See metadata.json for provenance and measured runtime.

Usage

from datasets import load_dataset
import numpy as np

data = load_dataset("fracapuano/reacher_context_split", split="train")
a = data.filter(lambda row: row["angle"] == "narrow")
b = data.filter(lambda row: row["angle"] == "broad")
z_a = np.asarray(a["embedding"], dtype=np.float32)
q2_a = np.asarray(a["q2"], dtype=np.float32)