datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
synthetic-dataset-1208
Synthetic Key-Value Retrieval 32K
This is a deterministic synthetic benchmark for exact key-value retrieval from
a long context. It is designed for evaluating long-context inference and KV
cache compression methods.
Context format
The context contains a one-time task description followed by an array:
You are given an array of key-value entries. Every key begins with K_ and every value begins with V_. Each entry has the format [key: value]. Given a query key, find… See the full description on the dataset page: https://huggingface.co/datasets/ollamaweights/synthetic-dataset-1208.synthetic-dataset-1208-64k
Synthetic Key-Value Retrieval 64K
This is a deterministic synthetic benchmark for exact key-value retrieval from
a long context. It is designed for evaluating long-context inference and KV
cache compression methods.
Context format
The context contains a one-time task description followed by an array:
You are given an array of key-value entries. Every key begins with K_ and every value begins with V_. Each entry has the format [key: value]. Given a query key, find… See the full description on the dataset page: https://huggingface.co/datasets/ollamaweights/synthetic-dataset-1208-64k.synthetic-kv-qwen3-8b-with-metadata
Synthetic KV Qwen3 8B — metadata-enhanced 64K
This dataset is an exact key-value retrieval benchmark. The context begins
with a short schema and task description, followed by records in the form
[KEY: VALUE]. Each question asks for the value belonging to one exact key.
The context is intentionally stored once in compact JSONL format. The
questions[i] entry corresponds to answers[i].
synthetic-kv-qwen3-8b