kaiwen2/wod-e2e-fast-ddrive-sasd-50k

WOD-E2E Fast-dDrive SASD 50k — TPU-ready training shards

50,331 frames (32 of the 263 WOD-E2E train shards) pre-tokenized into the Fast-dDrive Section-Aware Structured Diffusion (SASD) format, packaged as 787 Apache Parquet shards (~22 GB) for multi-host TPU training (grain / datasets / MaxText hf data path). The TPU side needs no tokenizer, processor, or torch — it loads arrays and trains.

⚠️ PRIVATE / license-restricted. Derived from the Waymo Open Dataset; shared privately under the WOD License (no redistribution). Do not make public.

Honest limitations

• —Pseudo text labels: trajectory + future_meta_behavior derive from real GT / ego-intent; critical_objects + explanation are heuristic pseudo-labels (WOD-E2E ships no text targets).
• —Subset: 50k of ~420k train frames. The full split runs through the identical pipeline (jax_ddrive/scripts/convert_subset_parallel.py --shards 263).
• —Uniform shapes: L=1184, nblocks=7, pixelvalues (672, 1176) — trivial batching.

Schema & reconstruction

Identical to kaiwen2/wod-e2e-fast-ddrive-sasd — see dataset_info_train.json for array_dtypes. Each array column is a little-endian binary blob + a <name>_shape list:

import numpy as np, pyarrow.parquet as pq, json
ADT = json.load(open("dataset_info_train.json"))["array_dtypes"]
row = pq.read_table("train-00000-of-00787.parquet").to_pylist()[0]
dec = lambda n: np.frombuffer(row[n], np.dtype(ADT[n])).reshape(row[n+"_shape"])
input_ids, pixel_values = dec("input_ids"), dec("pixel_values")

Columns: input_ids[L] i64, labels[L] i64, rbi[L] i32, turn[L] i32, scaffold[L] bool, weight_vec[L] f32, block_alpha/beta[n_blocks] f32, position_ids[3,L] i32, vision_mask[L] bool, pixel_values[N,1176] f16, image_grid_thw[n_img,3] i64.

Provenance

3-stage pipeline (autovla convert_wod_e2e.py → ddrive prep_train_jax.py → prep_to_parquet.py), orchestrated by convert_subset_parallel.py. Model: Efficient-Large-Model/Fast-dDrive (Qwen2.5-VL-3B block-diffusion VLA).