AIcell/guava-v13b

Note on this Hugging Face copy. The files below are the unmodified guava_v13b_data handoff package. Only this README.md was amended, with the YAML header above and this note, so that it renders as a dataset card — meaning its SHA-256 in checksums.json will not match. All other 15,404 entries in checksums.json were verified byte-for-byte before upload. The dataset viewer is disabled because training.jsonl holds nested chat messages with repo-relative image paths that the viewer cannot resolve.

Guava v13b — data-only training handoff

2,268 episodes: 1,554 main / 714 branch across 16 tasks.

Where to put the data

Copy the included data/version-13b folder into the same location in the recipient’s Guava checkout. Image paths are relative to the Guava repository root, matching the existing Guava training launchers that change into that directory. Point the launcher’s DATASET setting to data/version-13b/training.jsonl (or its absolute file path). Keep the included images directory in place; moving only the JSONL will not work.

Contents

• data/version-13b/training.jsonl: the ms-swift messages-format training export. Train on this file only.
• data/version-13b/v13_normalized.jsonl: matching canonical conversation-format data for inspection; not an additional training set.
• data/version-13b/images/: all required PNG images, copied without resizing or recompression. Identical encoded images share one file; episode/image order is unchanged.
• manifest.json, summary.json, image_index.json, provenance.json and validation.json: counts, provenance and integrity information.
• checksums.json at the package root: SHA-256 for every packaged file except this checksum file itself.
• No scripts, codebase, model weights, credentials, or training environment are included.

Data conventions and remaining checks

All message text, inline reasoning/tool calls, labels, loss flags and per-episode metadata are unchanged from v13b. Only the top-level runtime image paths are rewritten. Metadata retains historical provenance paths from the source machine; those are informational records, not files needed by the trainer. All runtime image references resolve within this package. Coordinates are already table-aligned: the tabletop is z = 0. Do not apply table-height normalization a second time. The source metadata retains calibration information. Release behavior is mixed across collection versions; retain the per-record prompts and feedback. The data has been through the recorded QC pipeline. The previously disclosed physical-clearance caveat for 05-24__push_cereal__trial_0308 remains; this packaging step did not run new simulator replays or remove further data. No held-out validation set is included. If creating a split, group parent/branch families together to avoid shared-history leakage. Check the effective split in the chosen training launcher. The recipient must check tokenization, assistant loss masking (including reasoning and tool calls), and long-episode handling in the actual Qwen training environment before a full run. Episodes may contain up to 30 images; the existing launcher’s context-length/image-token settings are not newly validated by this data handoff. Do not silently truncate or discard long episodes. No training has been started. Integrity verification here covers packaging, not a GPU training smoke test.

Per-task counts

Task	Main	Branch	Total
apple_juice_order	115	78	193
bread_near_lemon	160	24	184
can_in_bin	56	144	200
close_drawer	190	10	200
cube_stack	86	60	146
cube_under_cup	87	60	147
hotdog_near_donut	38	23	61
milk_near_cup	31	11	42
open_drawer	200	0	200
pick_up_orange	63	71	134
push_basket	75	7	82
push_cereal	114	8	122
red_objects_in_basket	134	66	200
remove_cube_from_tray	108	92	200
shell_game	54	19	73
tomato_in_bowl	43	41	84
Total	1,554	714	2,268