Long-context capabilities are a current bottleneck for LLMs: models can store tokens but struggle to align long-range reasoning and reward signals. This dataset supplies RL-ready training samples and diverse reward definitions to help bridge that gap—enabling experiments that pair long-context prompts with structured, task-specific reward functions.
What Sets It Apart
- Reward-focused structure: samples come with nine distinct reward/task categories (e.g., retrieval-style F1/NDCG, multiple-choice, summarization-style ground truth). So what: you can train or compare different RL objectives (policy-gradient, reweighting methods) on the same data distribution without rebuilding reward labels.
- Long-context orientation: items are constructed for long-document reasoning and evaluation on long-context benchmarks. So what: it reduces engineering overhead when evaluating long-range QA, multi-document summarization, and retrieval-aligned reward learning.
- Bilingual ground truth and metadata: each sample records language and structured ground-truth formats. So what: supports experiments that compare or mix Chinese and English long-context behaviors under identical reward formulations.
- Ready-to-load Parquet + datasets API compatibility: formatted for pragmatic use with common data stacks (datasets, polars, dask). So what: shortens iteration cycles for RL training pipelines and large-batch preprocessing.
Who It's For and Trade-offs
Great fit if you are researching or engineering RL for LLM long-context alignment—particularly teams benchmarking reward formulations, reweighting strategies, or RL algorithms (GRPO, AEPO, TMN‑reweight) for long inputs. Look elsewhere if you need massive-scale pretraining corpora, multimodal long-context data, or highly curated human preference labels: this dataset focuses on structured reward signals rather than large-scale raw web pretraining or high-volume human preference annotations.
Where It Fits
Use this dataset as the RL-training component alongside a model checkpoint when your goal is to improve long-context task-level performance (e.g., DocMath, long QA, multi-document summarization). It complements—rather than replaces—large supervised corpora or dedicated human preference datasets used for general instruction alignment.
Data & Reward Structure (brief)
Primary fields include prompt (chat-format messages), ability (task/reward category), and a reward_model block with ground_truth in list/string forms plus language metadata. That design makes it straightforward to compute both ranking-style metrics (NDCG/F1) and classification-style rewards during RL fine-tuning.
