tinydark-gram-data / README.md
hbfreed's picture
Upload README.md with huggingface_hub
6cfdb94 verified
|
Raw History Blame Contribute Delete
836 Bytes
---
license: cdla-sharing-1.0
---
# tinydark-gram-data
The exact data behind [tinydark-gram](https://github.com/hbfreed/tinydark-gram): GRAM (gradient-routed
auxiliary modules) on the time-of-day backdoor from Tiny Dark Stories.
- `tok_grid/`: TinyStories V2 and 242,688 normal/spooky story pairs, tokenized with a 4096-token BPE
(`tokenizer.json`). `.bin` files are uint16 token ids, `.idx` files are int64 story offsets.
- `pairs_train.jsonl`, `pairs_eval.jsonl`: the story pairs as text, used to fit the spooky classifier.
This is an earlier, smaller set of pairs than the released
[tiny-dark-stories](https://huggingface.co/datasets/hbfreed/tiny-dark-stories) dataset. Use that one for
anything other than reproducing tinydark-gram.
```bash
uvx hf download hbfreed/tinydark-gram-data --repo-type dataset --local-dir data
```