| | --- |
| | license: mit |
| | language: |
| | - en |
| | tags: |
| | - chemistry |
| | - physics |
| | - math |
| | - biology |
| | - science |
| | pretty_name: open-rl |
| | size_categories: |
| | - n<1K |
| | task_categories: |
| | - question-answering |
| | --- |
| | |
| | # Open-RL |
| |
|
| | [](https://opensource.org/licenses/MIT) |
| | [](https://turing.com) |
| |
|
| | --- |
| |
|
| | ## Dataset Summary |
| |
|
| | This dataset contains **self-contained, verifiable, and unambiguous STEM reasoning problems** across Physics, Mathematics, Biology, and Chemistry. |
| |
|
| | Each problem: |
| |
|
| | * Requires multi-step reasoning |
| | * Involves symbolic manipulation and/or numerical computation |
| | * Has a deterministic, objectively verifiable final answer |
| |
|
| | The problems were evaluated against contemporary large language models. Observed pass rates indicate that the tasks are **non-trivial yet solvable**, placing them within reach of advanced models while still exposing meaningful reasoning gaps. |
| |
|
| | This makes the dataset particularly suitable for: |
| |
|
| | * Reinforcement learning (RL) fine-tuning |
| | * Reward modeling |
| | * Outcome-supervised training |
| | * Verifiable reasoning benchmarks |
| |
|
| | --- |
| |
|
| | ## Dataset Structure |
| |
|
| | | Field | Type | Description | |
| | | ----------------- | ------ | ----------------------------------------- | |
| | | `conversation_id` | string | Unique identifier for each QA pair. | |
| | | `domain` | string | Physics, Math, Chemistry, Biology. | |
| | | `sub-domain` | string | Specific discipline. | |
| | | `question` | string | STEM problem statement (LaTeX supported). | |
| | | `answer` | string | Deterministic ground-truth solution. | |
| |
|
| | --- |
| |
|
| | ## Example |
| |
|
| | ```json |
| | { |
| | "conversation_id": "217998", |
| | "domain": "Physics", |
| | "sub-domain": "Astrophysics", |
| | "question": "Consider a Navarro–Frenk–White (NFW) dark matter halo profile where...", |
| | "answer": "\( \frac{4GM_{0}}{r_{0}} + \frac{16\pi Gk}{r_{0}}\left[ \ln\left(\frac{r_{0}}{r_{s}}\right) + 0.31 \right] \)" |
| | } |
| | ``` |
| |
|
| | --- |
| |
|
| | ## Verifiability and Automatic Grading |
| |
|
| | A core design principle of this dataset is **objective verifiability**. |
| |
|
| | Each problem is constructed such that: |
| |
|
| | * The final answer is deterministic |
| | * Correctness can be evaluated programmatically |
| | * No subjective interpretation is required |
| | * There is a clear separation between reasoning steps and final outcome |
| |
|
| | ### Answer Types |
| |
|
| | The dataset includes answers that are: |
| |
|
| | * Closed-form symbolic expressions |
| | * Numerical scalars |
| | * Algebraic identities |
| | * Simplified analytic forms |
| | * Canonical LaTeX representations |
| |
|
| | Because answers are deterministic, evaluation can be performed via: |
| |
|
| | * Exact string matching (after normalization) |
| | * Symbolic equivalence checking (e.g., SymPy) |
| | * Numerical tolerance comparison |
| | * Unit consistency validation (where applicable) |
| |
|
| | --- |
| |
|
| | ## Data Quality Assurance Process |
| |
|
| | To ensure scientific validity of the answer, all tasks are prepared and reviewed twice by PhD experts. |
| |
|
| | Key quality rubrics include: |
| |
|
| | * Prompt and answer accuracy |
| | * Clarity of prompt and underlying reasoning |
| | * Expert-verified model breaking cases due to model’s incorrect reasoning process |
| | * Google-proof originality validation. |
| |
|
| | --- |
| |
|
| | ## Reinforcement Learning and Outcome Supervision |
| |
|
| | This dataset is designed to support **outcome-based reinforcement learning** for reasoning models. |
| |
|
| | In contrast to preference-based RL (RLHF), which relies on subjective ranking signals, this dataset enables: |
| |
|
| | * Outcome-supervised reinforcement learning (OSRL) |
| | * Deterministic reward assignment |
| | * Binary or graded correctness rewards |
| | * Scalable automated evaluation |
| |
|
| | ### Example RL Setup |
| |
|
| | Given: |
| |
|
| | * Prompt: `question` |
| | * Model output: predicted final answer |
| |
|
| | Reward can be computed as: |
| |
|
| | * `+1` if the final answer matches ground truth |
| | * `0` or `-1` otherwise |
| | * Optional partial credit via symbolic or numerical closeness |
| |
|
| | This allows: |
| |
|
| | * Policy gradient methods (e.g., PPO) |
| | * Direct optimization against correctness signals |
| | * Reward model bootstrapping |
| | * Iterative self-improvement pipelines |
| |
|
| | ### Calibration Regime |
| |
|
| | The problems were stress-tested against advanced language models and found to be: |
| |
|
| | * Not trivially solved |
| | * Not universally failed |
| | * Within the capability frontier of modern LLMs |
| |
|
| | This places them in a **learning-efficient regime**: |
| |
|
| | * Hard enough to produce gradient signal |
| | * Solvable enough to avoid reward sparsity |
| | * Suitable for curriculum-style training |
| |
|
| | --- |
| |
|
| |
|
| | ## Future Directions: NuRL and Structured Nudging |
| |
|
| | We plan to extend this dataset with additional problem sets and a structured **"nudge" augmentation layer** inspired by the paper *["Nudging the Boundaries of LLM Reasoning"](https://arxiv.org/html/2509.25666v1)*. |
| |
|
| | ### Motivation |
| |
|
| | Standard online RL algorithms (e.g., GRPO-style approaches) can only learn from problems where the model occasionally produces correct rollouts. For sufficiently difficult problems with a **0% pass rate**, no reward signal is generated, and therefore no gradient updates occur. As a result, such problems cannot contribute to expanding the model’s reasoning frontier. |
| |
|
| | ### NuRL-Style Nudging |
| |
|
| | To address this limitation, future versions of this dataset will include: |
| |
|
| | * Abstract, high-level **hints ("nudges")** |
| | * Hints generated conditionally using the gold answer |
| | * Carefully designed cues that reduce problem difficulty without revealing the solution |
| |
|
| | Under a NuRL-style training pipeline: |
| |
|
| | 1. Rollouts are first generated without hints. |
| | 2. If pass rate > 0%, standard RL proceeds. |
| | 3. If pass rate = 0%, a structured hint is injected. |
| | 4. A new batch of trajectories is generated with the hint. |
| |
|
| | This enables: |
| |
|
| | * Previously unsolvable samples to produce non-zero rewards |
| | * Learning signal from frontier-level problems |
| | * Expansion of the model’s upper reasoning bound |
| |
|
| | ### Design Principles for Effective Nudges |
| |
|
| | Planned nudges will follow empirical findings from prior work: |
| |
|
| | * Hints should be **abstract and knowledge-oriented**, not answer-revealing |
| | * Hints should preserve distributional alignment with base policy reasoning |
| | * Hints should be injected only when necessary |
| | * Nudges are most effective after base RL convergence |
| |
|
| | --- |
| |
|
| | This evolution positions the dataset not only as a verifiable benchmark, but as a controlled testbed for **upper-bound expansion in reinforcement learning for reasoning models**. |
| |
|
| | --- |
| |
|
| | ## Citation |
| |
|
| | ```bibtex |
| | @dataset{turing_2026_open_rl, |
| | title = {Open-RL }, |
| | author = {Saurabh Patil, Anshuman Lall, Marko Pavlovic , Chinmayee Shukla, Seetesh Pande, Tejass Mohan Ukarde , Amanda Gollo Bertollo, Mahesh Joshi, Kihwan Han}, |
| | year = {2026}, |
| | url = {https://huggingface.co/datasets/TuringEnterprises/Open-RL} |
| | } |
| | ``` |
