jet-ai/ruler-100-nemotron
收藏资源简介:
RULER-100 — Nemotron-Nano-v3 tokenized数据集是RULER长上下文评估数据的重新生成版本,使用nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16(指令)分词器处理,以确保标注的上下文长度对该模型精确无误,避免使用其他模型(如Qwen3)分词数据时可能出现的漂移问题。数据集包含7种上下文长度(4096、8192、16384、32768、65536、131072、262144,即模型的最大长度)和13个RULER任务(niah_single_1/2/3、niah_multikey_1/2/3、niah_multivalue、niah_multiquery、vt、cwe、fwe、qa_1(SQuAD)、qa_2(HotpotQA)),每个(长度,任务)组合有100个样本,总计9100个示例。数据以原始提示形式存储(无聊天模板),生成时预留了16个令牌用于推理时的聊天包装器。数据布局为<length>/data/<task>/validation.jsonl,每个JSON行包含index、input、outputs和length字段。数据集由NVIDIA RULER(Apache-2.0许可证)生成,QA任务嵌入来自SQuAD v2(CC BY-SA 4.0)和HotpotQA(CC BY-SA 4.0)的段落,NIAH任务使用Paul Graham散文,生成种子为42。
RULER-100 — Nemotron-Nano-v3 tokenized is a regenerated version of RULER long-context evaluation data, tokenized with the nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 (instruct) tokenizer to ensure exact labeled context lengths for that model, avoiding drift that occurs when RULER data tokenized for a different model (e.g., Qwen3) is used with Nemotron. The dataset includes 7 context lengths (4096, 8192, 16384, 32768, 65536, 131072, 262144, the models maximum) and 13 RULER tasks (niah_single_1/2/3, niah_multikey_1/2/3, niah_multivalue, niah_multiquery, vt, cwe, fwe, qa_1 (SQuAD), qa_2 (HotpotQA)), with 100 samples per (length, task) combination, totaling 9,100 examples. Data is stored as raw prompts (no chat template), with 16 tokens reserved during generation for the chat wrapper applied at inference time. The layout is <length>/data/<task>/validation.jsonl, and each JSON line contains fields: index, input, outputs, and length. The dataset is generated by NVIDIA RULER (Apache-2.0 license), with QA tasks embedding passages from SQuAD v2 (CC BY-SA 4.0) and HotpotQA (CC BY-SA 4.0), NIAH tasks using Paul Graham essays, and a generation seed of 42.




