Datasets:

justintiensmith
/

VLA_Benchmark_Prompted_1200

license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - VLA
  - robotics
configs:
  - config_name: default
    data_files: data/*/*.parquet

VLA Benchmark Prompted 1200

This is a non-destructive LeRobot v3.0 merge of six 200-episode benchmark datasets. It contains 1,200 training episodes and 609,553 frames. Every episode's task string is the assigned prompt from meta/benchmark/prompt_manifest_v3.csv.

The Coke and Pepsi objects are canonically labeled as a crushed Coke can and a crushed Pepsi can. All rows are part of the training split.

Active prompt design

Prompt manifest v3 applies the exact 1,200 training prompts from VLA_Reasoning_Dataset_Info_Planned_After_61_Recollections.xlsx.

All 1,200 prompts exactly match the supplied final manifest.
State Recognition uses 49 explicit open/upside-down prompts and 151 category-only masked prompts.
closed, upright, and ambiguous either language remain withheld from State Recognition training prompts.
The other five task families retain the planned semantic and compositional masking rules from the supplied workbook.

The active machine-readable files are meta/benchmark/prompt_manifest_v3.csv and meta/benchmark/prompt_reassignment_audit_v3.csv.

State-recognition recollection overlay (40 episodes; historical)

On 2026-07-29, 40 checklist-selected episode slots were replaced, in checklist row order, with episodes 0–39 from justintiensmith/MT_State_Recognition_40 at 33c4b378b5867aa82ef7fd0b4f6157f0c11f7a05. Existing episode prompts and task indices were preserved. The resulting dataset contains 614,367 frames.

State-recognition recollection overlay (61 episodes)

On 2026-07-30, 61 additional checklist-selected episode slots were replaced, in checklist row order, with episodes 0–60 from justintiensmith/SP_State_Recognition_61 at 0a847ce8031c22a19e5adcbcc44d63b4d6cbd4dd. Together with the prior 40-episode overlay, 101 State Recognition demonstrations have now been recollected. The resulting dataset contains 609,553 frames and uses the v3 prompt manifest.

The source workbooks are retained under meta/benchmark/.

Immutable source revisions

justintiensmith/MT_State_Recognition_200 at 168e12e735da9ad62716dca89cff660b53082cc3
justintiensmith/SP_Relational_Placement_200 at 8eff46af18af7a9a6f763ed975f8b3357c673685
justintiensmith/SP_Referential_Disambiguation_200 at 2642784f74a0c91bc118702b7993895aacd311f5
justintiensmith/SP_Sequencing_200 at 1cda4c4c3c6ae1fa4b5dc6d2cf4c887bb9f1da4e
justintiensmith/SP_Counting_200 at 6591c0f19952e23010b7c6a0630e4d60f8714cbb
justintiensmith/MT_Size_Recognition_200 at 3a6e75d91e8339a87eebd8e545d87b148ae19925

Local identifier

justintiensmith/VLA_Benchmark_Prompted_1200

Episode 819 startup trim

Episode 819 (Put the blue pen in the bowl.) is unchanged after its initial 0.4 seconds. Twelve startup frames were removed to exclude one invalid wrist_flex.pos state reading at 0.2 seconds. The shared video payloads were not re-encoded; all five episode video start offsets advance by 0.4 seconds. The tabular data, indices, numeric episode statistics, and global statistics were rebuilt after the trim.