---
license: apache-2.0
pretty_name: sortletters serialized 121 unqualified engineering dataset
task_categories:
- robotics
tags:
- LeRobot
- engineering-fixture
- rejected-for-behavior-training
configs:
- config_name: default
  data_files: data/*/*.parquet
---

# sortletters_serialized_121

> **STOP — not qualified as a coherent behavior-training baseline.** The files are internally
> indexed and decodable, but the available provenance does not establish a valid real-robot
> sortletters policy-training dataset or success rate.

> **警告：该数据未通过一致行为训练基线资格门。** 文件可读取、索引一致，但不能据此声称
> 有效真机字母排序策略或真机成功率。

Read-only source audit findings for revision
`344ef8e12b2f9f5e5bc4c25be042dde085cabb12`:

- No persisted per-episode success labels or autonomous success-verifier records exist.
- Episodes 84–89 contain visible human presence and direct object intervention during the task.
- All 121 exact terminal right-front frames visually show `M I N T`, while the only persisted
  prompt is `字母排序`; it does not specify MINT and naturally conflicts with alphabetical
  `I M N T`.
- All 121 exact start frames use one fixed letter permutation; no letter-permutation diversity
  was observed.
- Wrist videos contain repeated decoded frames during associated-arm motion, and original
  per-sensor capture timestamps were not recorded.

At most 115 episodes remain a candidate for a separately named, fixed-layout
`MINT`-rearrangement dataset after explicit semantic relabeling and additional trajectory review.
This card does not certify those episodes as successes.

Allowed use without further qualification: dataset inspection, loader tests, and runtime engineering.