strifero/attribution-groundtruth

54 downloads in 30 days0 likesLast modified 27 Oct 2026View on Hugging Face

The declared licence (apache-2.0) is classified as permissive.

Declared licence: apache-2.0, from the tagEvidence: README.md at 02fc811
Licence files: No licence file at the top level of the repositoryEvidence: file list at 02fc811
Declared sources: None listed in the card metadata
Access gated on Hugging Face: No

Each cell links to what was checked: a file, a file list or a Hub record.

Citation text from the card

A benchmark for methods that claim to tell you which training data taught a model something. Two small models were trained from scratch on the same ordered 5B-token stream, differing only in which half of 528 fictional facts their planted documents stated. Because we planted the documents, we know the right answer; because we have the twin that never saw each fact, we can measure what each document actually did.

- Models: [tda-124m-A](https://huggingface.co/strifero/tda-124m-A) and [tda-124m-B](https://huggingface.co/strifero/tda-124m-B), with checkpoints every 0.5B tokens
- Report: [Paper trail](https://trifero.ai/research/paper-trail/)
- Code (stream builder, training, every method, scoring): [https://github.com/strifero/attribution-groundtruth](https://github.com/strifero/attribution-groundtruth)

All planted content is fictional. The base training text is not redistributed here; `stream/` holds the exact recipe and order to rebuild it, with a hash of every window to check the rebuild.

All datasets marked Permissive or share-alike | JSON

The same facts for your agent

A free MCP server at mcp.datalicense.uk/mcp. Connect it from an MCP client such as Claude or Cursor.

dataset_licence({ "id": "strifero/attribution-groundtruth" })
signal: "permissive"
stated_licence: "apache-2.0"
sources: 0