mmlu / README.md
JonathanMiddleton's picture
Update README.md
a840089 verified
# mmlu (filtered)
Filtered version of [cais/mmlu](https://huggingface.co/datasets/cais/mmlu).
## Source Dataset
- **Original HuggingFace path**: `cais/mmlu`
- **Disjoint train/val splits**: yes
- **train split**: 1 subset(s) from `train` split
- **val split**: 57 subset(s) from `test` split
## Filtering Methodology
Each example in the original dataset was reviewed by an LLM (GPT-OSS-120B) which assessed quality across several dimensions including factual accuracy, formatting consistency, and instructional clarity. Examples that failed review were collected into a blocklist keyed by source name and original row index.
The filtered dataset was produced by:
1. Loading the original dataset from HuggingFace
2. Stamping each row with its original index (`__source_orig_idx__`) for traceability
3. Removing all rows whose index appears in the blocklist
4. Validating that row counts match expectations exactly
5. Spot-checking a random sample of surviving rows against the original to verify data integrity
## Filtering Impact
| Split | Original Rows | Removed | Filtered Rows | % Removed |
|-------|--------------|---------|---------------|-----------|
| train | 99,842 | 6,928 | 92,914 | 6.94% |
| val | 14,042 | 0 | 14,042 | 0.00% |
## Schema
All original columns are preserved. One column is added:
- `__source_orig_idx__`: The row's index in the original (unfiltered) dataset. This provides complete lineage back to the source for debugging and future analysis.
**Note**: The original `train` split had a nested structure (`subsplit_name: "train"`). This has been flattened into top-level columns. When updating `data_source_configurations.py` to point to this repo, remove the `subsplit_name` field from the `train` split config.
## Provenance
- **Blocklist**: `data/instruct_mix/train/review_results/blocklist_proposed_20260218_232356.jsonl`
- **Mixture**: `assistant_core_midtraining`
- **Generated**: 2026-02-20T12:05:03.203204+00:00