jungter/potao-gang-gc-032026

---
license: mit
task_categories:
- text-generation
language:
- en
pretty_name: Not Potato Gang
---

_new is cleaned using a varity of methods like duplicate message removal, then fed to an LLM to go through A and B 0 - n classifying if they are related pairs or not. The result model is [here](https://huggingface.co/jungter/potato-gang-gc)

_v3 is cleaned with the same algo, however the LLM is fed 7 messages plus 20 context padding before and after and asked to pair prompt and responses, with 3 messages overlap between chunks. The result model is [here](https://huggingface.co/jungter/schizobot9000)

With the same training params, _v3 yielded a more talkitive and coherent model, however is prone to repetition and rambling on about topics. In contrast, _new yielded a "dumber" model but is more grounded.

The LLM used for classification is [Qwen/Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B)