26 lines
1.8 KiB
Markdown
26 lines
1.8 KiB
Markdown
---
|
|
language:
|
|
- en
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- qwen3
|
|
- hallucination-detection
|
|
- sopo
|
|
library_name: transformers
|
|
---
|
|
|
|
# SOPO-4B
|
|
|
|
Official release of the 4B [SOPO (Socratic Policy Optimization)](https://github.com/patsnap/SOPO) model for span-level hallucination detection. This model is presented in the paper [SOPO: Socratic Guided Policy Optimization for Span-Level Hallucination Detection](https://huggingface.co/papers/2610.32212). See the [evaluation instructions](https://github.com/patsnap/SOPO#evaluation) for the RAGTruth test set.
|
|
|
|
## RoPE configuration and compatibility
|
|
|
|
**This checkpoint requires a RoPE base (`rope_theta`) of `5000000`, not the default `10000`.** The original `config.json` saved with Transformers 5.8.1 contained this value only under `rope_parameters.rope_theta`. With some older inference stacks, including Transformers 4.57.1 and vLLM 0.14.0, the absent top-level `rope_theta` can default to `10000` and override the nested value without an error or warning. This can seriously degrade predictions and cause unusually long, token-limited generations.
|
|
|
|
The updated `config.json` sets **both** `rope_theta` and `rope_parameters.rope_theta` to `5000000` for compatibility. The model weights are unchanged. If you downloaded the checkpoint before this correction, re-download `config.json` (or add `"rope_theta": 5000000` at the top level of your local copy) before running inference. Check that the effective RoPE base in your serving stack is `5000000`, especially when using older Transformers or vLLM versions.
|
|
|
|
## Related resources
|
|
|
|
- [SOPO source code and evaluation script](https://github.com/patsnap/SOPO)
|
|
- [RL training dataset](https://huggingface.co/datasets/OverwhelmingFit/SOPO-RL-RAGTruth10k)
|
|
- [Reasoning-augmented SFT training dataset](https://huggingface.co/datasets/OverwhelmingFit/SOPO-SFT-RAGTruth5k-Reasoning) |