jungter/potao-gang-gc-032026
_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 _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 With the same training params, _v3 yielded a more talkitive and coherent model, however is… See the full description on the dataset page: https://huggingface.co/datasets/jungter/potao-gang-gc-032026.
_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
_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
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