kaiwen2/wod-e2e-fast-ddrive-sasd

WOD-E2E Fast-dDrive SASD — TPU-ready training shards

Pre-tokenized Section-Aware Structured Diffusion (SASD) training samples for the Fast-dDrive Qwen2.5-VL-3B block-diffusion driving model, packaged as sharded Apache Parquet for multi-host TPU training (grain / datasets / MaxText hf data path).

Each row is one Waymo Open Dataset End-to-End (WOD-E2E) front-camera frame, already run through the Fast-dDrive prep (chat template + Qwen2.5-VL processor + deep-JSON scaffold + 3D M-RoPE positions + per-section block indices), so the TPU side needs no tokenizer, no processor, no torch — it just loads arrays and trains.

⚠️ PRIVATE / license-restricted. This data is derived from the Waymo Open Dataset and is shared privately under the WOD License, which prohibits redistribution. Do not make this dataset public or share it outside your licensed use.

Honest limitations (read before using)

• Pseudo text labels. WOD-E2E ships no native text targets. trajectory and future_meta_behavior derive from real GT / ego-intent; critical_objects and explanation are heuristic pseudo-labels. Train/eval accordingly.
• Subset. This is a ~400-frame representative subset. The full 263-shard WOD-E2E train split runs through the identical 3-stage pipeline (see "Provenance").
• Not TPU-verified by the author. The packaging + decode is bit-exact verified on CPU/GPU; end-to-end TPU training is the intended downstream use, validated on emulation + a single GPU.

Schema (per row)

Scalars: sample_id (str), L (int32), n_blocks (int32). Arrays — each a little-endian binary blob + a <name>_shape int list:

column	dtype	shape	meaning
input_ids	int64	[L]	token ids (mask-padded to multiple of bd_size=32)
labels	int64	[L]	-100 except assistant-response tokens
rbi	int32	[L]	response-block index per token (-1 = none)
turn	int32	[L]	turn index for the hybrid block-causal mask
scaffold	bool	[L]	fixed JSON-scaffold positions (never noised)
weight_vec	float32	[L]	per-token section-importance weight
block_alpha / block_beta	float32	[n_blocks]	per-block Beta noise schedule
position_ids	int32	[3, L]	3D M-RoPE positions
vision_mask	bool	[L]	image/vision-start token positions
pixel_values	float16	[N, 1176]	Qwen2.5-VL patch features (3 front cams)
image_grid_thw	int64	[n_img, 3]	per-image temporal/height/width grid

Reconstruct (numpy, framework-free)

import numpy as np, pyarrow.parquet as pq
ARRAY_DTYPES = {  # see dataset_info_train.json
  "input_ids":"int64","labels":"int64","rbi":"int32","turn":"int32","scaffold":"bool",
  "weight_vec":"float32","block_alpha":"float32","block_beta":"float32",
  "position_ids":"int32","vision_mask":"bool","pixel_values":"float16","image_grid_thw":"int64"}
row = pq.read_table("train-00000-of-00007.parquet").to_pylist()[0]
def dec(name): return np.frombuffer(row[name], np.dtype(ARRAY_DTYPES[name])).reshape(row[name+"_shape"])
input_ids = dec("input_ids"); pixel_values = dec("pixel_values")

See load_example.py for a runnable version.

Provenance (3-stage pipeline, reproducible at full scale)

1. fast_ddrive/data/convert_wod_e2e.py (autovla env): tfrecord → train JSON + JPEGs (--with_target).
2. jax_ddrive/eval/prep_train_jax.py (ddrive env): JSON+JPEGs → per-sample SASD npz.
3. jax_ddrive/ddrive_jax/convert/prep_to_parquet.py → these Parquet shards.

Model: Efficient-Large-Model/Fast-dDrive (Qwen2.5-VL-3B block-diffusion VLA).