| | --- |
| | license: other |
| | datasets: |
| | - LightningRodLabs/future-as-label-paper-training-dataset |
| | base_model: |
| | - Qwen/Qwen3-32B |
| | pipeline_tag: reinforcement-learning |
| | --- |
| | # Future-as-Label: Scalable Supervision from Real World Outcomes |
| |
|
| |  |
| |
|
| | - **Paper:** https://arxiv.org/abs/2601.06336 |
| | - **Data generation platform:** https://lightningrod.ai |
| | - **Training dataset:** https://huggingface.co/datasets/LightningRodLabs/future-as-label-paper-training-dataset |
| |
|
| | ## Highlights |
| |
|
| | This model is a forecasting-specialized language model trained using Future-as-Label, an outcome-based supervision and data generation approach that leverages the natural resolution of real-world events over time. Instead of relying on human annotation or proxy objectives, supervision is provided entirely by externally verifiable outcomes once events resolve. |
| |
|
| | Built as a fine-tuned derivative of **Qwen3-32B**, the model is optimized for probabilistic prediction under uncertainty. Its training objective emphasizes calibration, decision quality, and outcome-aligned learning rather than static target matching. |
| |
|
| | Key highlights include: |
| |
|
| | - **Outcome-Based Supervision at Scale** Training rewards are computed retrospectively using proper scoring rules once real-world events resolve, enabling scalable supervision without manual annotation. |
| | - **Causal and Temporal Constraints** Predictions are generated using only information available at prediction time, while evaluation is deferred until outcome resolution, enforcing strict causal separation. |
| | - **Improved Calibration and Accuracy** Relative to its pretrained baseline, the model achieves substantially lower Brier scores and expected calibration error, and outperforms a much larger same-generation model (Qwen3-235B) on real-world forecasting benchmarks. |
| |
|
| | ## Model Description |
| |
|
| | This model is a decoder-only large language model fine-tuned from Qwen3-32B using LightningRod Labs’ Future-as-Label training methodology, also referred to as Foresight Learning. |
| |
|
| | The model is trained to produce probabilistic forecasts about real-world events from causally masked inputs. Rather than fitting fixed targets via supervised fine-tuning, learning is driven entirely by outcome-based rewards computed after events resolve using proper scoring rules. This aligns optimization directly with calibration and predictive accuracy under uncertainty. |
| |
|
| | While current experiments focus on binary outcomes, the underlying framework generalizes naturally to multi-class, continuous, and richer outcome spaces. |
| |
|
| | ## Model Sources |
| |
|
| | - **Developed by:** LightningRod Labs |
| | - **Model type:** Large language model (decoder-only transformer) |
| | - **Language(s):** English (primarily) |
| | - **License:** Qwen3 License |
| | - **Finetuned from:** Qwen/Qwen3-32B |
| |
|
| | ## Uses |
| |
|
| | ### Direct Use |
| |
|
| | Research and applied experimentation involving: |
| | - Binary event prediction and probabilistic forecasting |
| | - Real-world outcome modeling with delayed feedback |
| | - Studies of calibration, uncertainty estimation, and decision quality |
| | - Synthetic data generation for forecasting and evaluation pipelines |
| |
|
| | ### Out-of-Scope Use |
| |
|
| | The model is not intended for: |
| | - Safety-critical or regulated domains |
| | - Deployment without additional evaluation |
| | - Use cases restricted by the Qwen3 license terms |
| |
|
| |
|
| | ## Limitations |
| |
|
| | As a fine-tuned derivative model, behavior may differ from the base Qwen3 model and may exhibit hallucinations or reasoning errors. |
| |
|
| | ## License and Attribution |
| |
|
| | This model is a fine-tuned derivative of **Qwen3-32B**. |
| |
|
| | The model weights are released under the **Qwen3 License**. |
| | All original license terms, conditions, and attribution requirements apply. |
| |
|
| | See the original Qwen3 license for full details. |
| |
|
