Datasets:
Tasks:
Question Answering
Modalities:
Text
Formats:
json
Languages:
English
Size:
10K - 100K
ArXiv:
License:
|
Download README.md from speediedan/knowing-using-gap-chaining: direct link, hf CLI and curl.
- Browser
- Download file 2.53 kB
-
https://huggingface.co/datasets/speediedan/knowing-using-gap-chaining/resolve/main/README.md
- Command line
-
hf download hf://datasets/speediedan/knowing-using-gap-chaining/README.md
-
curl -L -o README.md https://huggingface.co/datasets/speediedan/knowing-using-gap-chaining/resolve/main/README.md
2.53 kB
| license: mit | |
| language: | |
| - en | |
| pretty_name: Knowing-Using Gap chaining tasks (STaRK-Prime) | |
| task_categories: | |
| - question-answering | |
| tags: | |
| - knowledge-injection | |
| - multi-hop | |
| - biomedical | |
| size_categories: | |
| - 10K<n<100K | |
| # Knowing-Using Gap: chaining tasks | |
| A durable copy of the two-hop chaining data from [Towards Mechanistically Understanding Why Memorized Knowledge Fails | |
| to Generalize in Large Language Model Finetuning](https://www.alphaxiv.org/abs/2607.08393) (Dai et al.), published so | |
| that reproductions do not depend on the paper's anonymous review mirror, which expires in March 2027. The file is the | |
| authors' `data/multi_fact_data_template/chaining_tasks.json`, unchanged. | |
| ## Contents | |
| `chaining_tasks.json`: 11,997 items built from the STaRK-Prime biomedical knowledge base, 1,333 for each of 9 relation | |
| paths (for example anatomy, expression present, gene or protein, target, drug). Each item holds: | |
| - `facts`: two single-hop facts (`head`, `relation`, `tail`) whose composition is the item's question; | |
| - `memorization_tasks`: one question and answer per fact, the text a model is fine-tuned on; | |
| - `generalization_tasks`: the two-hop question that composes the two facts, and its answer; | |
| - `task_type`: `chaining`. | |
| ## Use | |
| The paper injects 1,000 facts (500 items) and scores greedy exact match on the two-hop questions. Its sample is drawn | |
| with Python's `random`: seed the generator (848 for Qwen models, 1340 for Llama models), shuffle the items with their | |
| indices, then add items in that order to a randomized subset sum over their memorization-task counts until exactly | |
| 1,000 tasks are reached. Drawn this way, the Qwen sample has 409 unique two-hop questions and the Llama sample 427. | |
| Interpretune's `knowing_using` example implements the draw and reproduces the paper's LRSD result on Qwen2.5-1.5B. | |
| ## Provenance and license | |
| - Source: the paper's code and data release (MIT License, copyright 2026 The Mem2Gen Authors; `LICENSE` here is its | |
| text), retrieved from `https://anonymous.4open.science/r/Mem2Gen-71FF`. | |
| - `chaining_tasks.json` sha256: `7ea065e4e3ddaa0f6ee100695961ddf8f9a9581938e5b6127168858718598d5d`. | |
| - The items derive from [STaRK](https://github.com/snap-stanford/stark) (MIT License), whose STaRK-Prime knowledge base | |
| builds on [PrimeKG](https://github.com/mims-harvard/PrimeKG). As the paper's authors ask, please also respect the | |
| terms of those upstream resources. | |
| ## Citation | |
| Cite the paper (Dai et al., arXiv 2607.08393), and STaRK and PrimeKG for the underlying knowledge base. | |