MBT-Validation / README.md
gaotang's picture
Release fixed MBT validation sequences
bd70808 verified
|
Raw History Blame Contribute Delete
2.49 kB
---
pretty_name: MBT Validation
license: odc-by
language:
- en
size_categories:
- 10K<n<100K
task_categories:
- text-generation
tags:
- tokenized
- language-model-evaluation
- multi-branch-transformer
source_datasets:
- HuggingFaceFW/fineweb-edu
configs:
- config_name: default
data_files:
- split: validation
path: data/validation-00000-of-00001.parquet
---
# MBT Validation
The fixed validation sequences used for [Multi-Branch Transformer](https://arxiv.org/abs/2611.19800). This release preserves the original token IDs and sample order, allowing evaluation without rebuilding dataset indexes or downloading the full corpus.
| Split | Rows | Column | Tokens per row | Evaluated targets |
|---|---:|---|---:|---:|
| validation | 12,160 | `input_ids`: list of int32 | 4,097 | 49,807,360 |
Each row contains 4,096 input tokens plus one additional next-token target. All targets contribute to the loss, so no loss-mask column is needed. Preserve row order and token IDs; do not retokenize or add special tokens.
```python
from datasets import load_dataset
data = load_dataset("gaotang/MBT-Validation", split="validation")
row = data[0]["input_ids"]
inputs = row[:-1]
targets = row[1:]
```
Use causal attention and positions `0` through `4095` within each sequence. Neither positions nor attention reset at document boundaries. Compute validation loss as the sum of next-token cross-entropies divided by the number of evaluated targets; perplexity is its exponential.
## Source and preparation
The source is [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu), configuration `sample-100BT`, at revision `87f09149ef4734204d70ed1d046ddc9ca3f2b8f9`. We held out [sample/100BT/000_00000.parquet](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu/blob/87f09149ef4734204d70ed1d046ddc9ca3f2b8f9/sample/100BT/000_00000.parquet) from MBT pretraining. This is our held-out subset, not an official upstream validation split.
Documents were tokenized with the Llama 3 tokenizer, ended with EOD token `128001`, and packed into sequences. The release contains the first 12,160 samples in the original validation order generated with seed `1234`. The vocabulary contains 128,256 token IDs; tokenizer files are not included.
FineWeb-Edu is provided by Hugging Face under [ODC-By](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu/blob/87f09149ef4734204d70ed1d046ddc9ca3f2b8f9/README.md). This derived dataset retains that attribution and license.