acvlab/ABot-ER-Training-Data

ABot-ER Training Data

ABot-ER Training Data is the JSONL training release used for the ABot-ER multi-capability foundation and final joint post-training stages. It covers spatial and relational reasoning, scene and state understanding, temporal and event reasoning, point grounding and affordance, and long-horizon action organization.

Dataset structure

Configuration (each has a train split)	Purpose	Physical records	Shards
stage_a_pretrain	Multi-capability foundation training	462,385	288
stage_b_posttrain	Final joint post-training	104,960	128

Files are stored under:

data/stage_a_pretrain/*.jsonl
data/stage_b_posttrain/*.jsonl

The release preserves the physical exposure order of the sealed training data. Stage B therefore contains intentional repeated exposures where present in the training release.

Download and load by stage

Install huggingface_hub to download the files. Set HF_REVISION to a fixed commit SHA from acvlab/ABot-ER-Training-Data for reproducibility.

import os
from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="acvlab/ABot-ER-Training-Data", repo_type="dataset",
    revision=os.environ["HF_REVISION"],
    allow_patterns=["data/stage_a_pretrain/*.jsonl", "data/stage_b_posttrain/*.jsonl",
                    "manifests/*", "*.md", "*.json", "*.txt", "*.bib", "LICENSE"],
)

The card declares two configurations: stage_a_pretrain and stage_b_posttrain. Both contain training data only; do not treat one stage as validation/test data. Media inventory records are not training examples and are excluded from both configs. The JSONL has task-dependent nested metadata. To preserve its exact values and physical order without Arrow schema coercion, use the standard-library reader:

import json
from pathlib import Path

def iter_stage(root, stage):
    if stage not in {"stage_a_pretrain", "stage_b_posttrain"}:
        raise ValueError(stage)
    for shard in sorted((Path(root) / "data" / stage).glob("*.jsonl")):
        with shard.open(encoding="utf-8") as f:
            for line in f:
                yield json.loads(line)

first = next(iter_stage(root, "stage_b_posttrain"))

After publication, HF datasets users can select either declared configuration:

from datasets import load_dataset

stage_a = load_dataset(
    "acvlab/ABot-ER-Training-Data", name="stage_a_pretrain", split="train",
    revision=os.environ["HF_REVISION"], streaming=True,
)
stage_b = load_dataset(
    "acvlab/ABot-ER-Training-Data", name="stage_b_posttrain", split="train",
    revision=os.environ["HF_REVISION"], streaming=True,
)

Local JSON streaming was verified with datasets==4.2.0 over all 462,385 Stage A and 104,960 Stage B records, with no record-count mismatch. The hosted HF viewer was not tested. If a different library version reports a nested type/cast error, use the standard-library reader above rather than dropping or silently changing records. Downloading annotations does not make the dataset immediately runnable: referenced images/videos must be acquired separately and mapped by the user. No data-construction pipeline is included in this release.

Record format

Each JSONL record contains:

• conversations: user/assistant training messages;
• images and videos: repository-relative media references;
• task_family: 2d_qa or task_plan;
• benchmark_proxy, capability_primary, answer_format, and coordinate metadata;
• canonical source name and repository;
• available question, media, trajectory, and grouping hashes.

Task-specific output contracts are retained in the assistant label. VABench-P uses eight normalized points wrapped in its XML answer contract; Where2Place point-supervision records use a tuple list. Auxiliary open-text records preserve their original answer type. Multiple-choice, numeric, open-text, and executable-plan records retain their benchmark-aligned schemas.

Task-Conditioned Adaptive Training Strategy

Optional route_token and route_source metadata describe the task-family routing. Route tokens are injected by the model data loader and are not physically added to the frozen user prompt:

• 2d_qa records use [2D_QA];
• task_plan records use [TASK_PLAN].

Models without task-conditioned residuals can train on the same JSONL without route-token injection.

Media

This repository is a JSONL-only release. Image and video binaries are not duplicated here. The images and videos fields are stable relative references, while manifests/media_manifest.jsonl provides hashes and source information for media acquisition and verification.

Sources and terms

The corpus combines public embodied vision-language datasets and ABot-ER annotations. Canonical sources, record counts, upstream repositories, and applicable terms are listed in manifests/source_manifest.csv. Source-specific terms continue to apply. See LICENSE and named licenses and source links. HANDAL- and HOI4D-derived records are restricted to non-commercial research use under their respective licenses.

Intended use

This release is intended for non-commercial technical research and academic use in embodied multimodal understanding and planning. Users are responsible for obtaining referenced media from the corresponding upstream sources and complying with source-specific terms.

Privacy and integrity

The public records contain no private NAS paths, personal workspace paths, internal model URIs, or credentials. File hashes are provided in SHA256SUMS.txt, and aggregate release statistics are recorded in RELEASE_MANIFEST.json.

Citation

ABot-ER: Unified Embodied Reasoning through Agentic Data-Model Co-Evolution.

@misc{ding2026aboterunifiedembodiedreasoning,
  title={ABot-ER: Unified Embodied Reasoning through Agentic Data-Model Co-Evolution},
  author={Qiuyu Ding and Heng-Da Xu and Changda Xia and Jiajun Shen and Zihao Guan and Zhengchen Yuan and Jianfang Zhang and Di Yang and Jiayi Tian and Yuting Xu and Fei Wang and Shiao Liu and Wei Zhang and Xiaolong Wu and Wenbin Tang and Zedong Chu and Mu Xu},
  year={2026},
  eprint={2610.18355},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2610.18355}
}