adamo1139/Experimental-DeepSeek-Coder-V2-Lite-Jump-alpha2-GGUF

DeepSeek-Coder-V2-Lite-Base finetuned for 1 epoch on adamo1139/JUMP_Coder_mini_v1-3 via llama-factory at 2500ctx with qlora, rank 40 and alpha 40.

Prompt format is deepseek-coder as indicated by chat_template in tokenizer_config.jsom.

First impressions: much better than alpha1, actually decent performance on Powershell, there's some weirdness around system prompt being output at the end of assistant reply.

It's an early WIP, unless you are dying to try DeepSeek-Coder-V2-Lite finetunes I suggest you don't use it :)