Add model card and metadata
#1
by nielsr HF Staff - opened
README.md
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: transformers
|
| 3 |
+
pipeline_tag: text-generation
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# PVminerLLM2_3B
|
| 7 |
+
|
| 8 |
+
PVminerLLM2 is a set of Large Language Models (LLMs) optimized for the structured extraction of patient voice information from unstructured text. This model specifically represents the 3B parameter variant.
|
| 9 |
+
|
| 10 |
+
The model was introduced in the paper [PVminerLLM2: Improving Structured Extraction of Patient Voice via Preference Optimization](https://huggingface.co/papers/2606.16074).
|
| 11 |
+
|
| 12 |
+
## Model Description
|
| 13 |
+
|
| 14 |
+
PVminerLLM2 improves upon prior work in structured patient voice extraction by applying preference optimization to address token-critical errors that supervised fine-tuning (SFT) often misses. These errors are frequently rare, fine-grained, and unevenly distributed in structured outputs.
|
| 15 |
+
|
| 16 |
+
Key methodological improvements include:
|
| 17 |
+
- **Token-level Gated Stabilization:** A preference objective term that prevents the degradation of absolute token likelihood.
|
| 18 |
+
- **Confusion-aware Preference Pair Construction:** A strategy to better capture low-separation distinctions between different extraction categories.
|
| 19 |
+
- **Reweighing Strategies:** Incorporation of token-importance weighting and inverse-frequency reweighing to address token imbalance and class skew.
|
| 20 |
+
|
| 21 |
+
Across various model sizes, PVminerLLM2 consistently outperforms standard baselines and models trained with existing preference optimization methods.
|
| 22 |
+
|
| 23 |
+
## Links
|
| 24 |
+
|
| 25 |
+
- **Paper:** [PVminerLLM2: Improving Structured Extraction of Patient Voice via Preference Optimization](https://huggingface.co/papers/2606.16074)
|
| 26 |
+
- **GitHub Repository:** [Data-Mining-Lab-Yale/PVminerLLM2](https://github.com/Data-Mining-Lab-Yale/PVminerLLM2)
|
| 27 |
+
|
| 28 |
+
## Implementation Details
|
| 29 |
+
|
| 30 |
+
The official repository provides a full training and evaluation pipeline, including scripts for generating confusion summaries, building preference data, training with LoRA adapters, and merging models. For detailed implementation instructions, please refer to the [GitHub README](https://github.com/Data-Mining-Lab-Yale/PVminerLLM2).
|