Published dataset is incompatible with the loader and all splits are identical

## Summary

The dataset linked from the LPCD project page cannot be consumed by the released training loader, and its six advertised train/validation/test CSV files are all the same Git object.

Repository baseline: `a0b47382f08b108a17a39b343c4e58dbf2852db3`.
Dataset checked: `ByteDance/LiveStreamingRiskControl` on 2026-08-22.

## Published artifact evidence

The Hugging Face tree API reports the following for all six files:

- `May_train.csv`, `May_val.csv`, `May_test.csv`
- `June_train.csv`, `June_val.csv`, `June_test.csv`

Each file is 3,062,039 bytes and has the identical Git blob OID:

```text
d945552c272b96d1e5eac6eb04b17eb75b052268
```

Their common header is:

```csv
room_idx,room_label,date
```

Therefore May/June and train/validation/test currently contain byte-identical rows rather than independent splits.

## Loader incompatibility

`dataset_preprocessed_new.py` requires CSV fields that are absent from the published files:

- line 40 reads `room_id` (the dataset publishes `room_idx`);
- lines 48 and 51 read `fraud_second_label` (not published).

It also hard-codes private `/mnt/bn/open/...` CSV and LMDB paths at lines 30–38. The referenced LMDB records containing `patch_list` are not part of the public dataset, so renaming one CSV column would still not make the training pipeline runnable.

A minimal DataFrame using the published header raises `KeyError: 'room_id'` at the first loader access; after substituting `room_idx`, it raises `KeyError: 'fraud_second_label'`.

## Expected behavior

- train, validation, and test artifacts should be distinct according to the documented split policy;
- the public schema and payload should match a configurable loader in this repository;
- the README/project page should document the exact command and paths needed to train from the public release.

## Impact

Using these files as named produces complete train/validation/test leakage, while the released code cannot initialize from them at all. This blocks reproduction and could produce invalid evaluation results if consumers assume the filenames represent real splits.

## Why there is no PR in this report

A complete fix requires republishing the correct dataset artifacts (including a decision about the missing LMDB feature payload and label fields) and defining the intended split policy. Those source data and product decisions are not available in this repository, so a code-only patch would be incomplete.

Start with dataset_preprocessed_new.py, especially lines 30–38 and 40, 48, and 51, then compare its required fields and paths with the published CSV header and missing LMDB payload. Review the README and project page references. Done means the public artifacts have distinct documented splits, the loader can be configured for their schema and payload, and training commands and paths are documented.

Written by the indexing model from the issue text.