A highly experimental, underfitted tiny language model based on Qwen3 base, then upgraded to use the instruct model's lm head and continued training afterwards.

gguf | safetensors

To create a model capable of roleplaying as a character, and also think and call tools to allow for interactive npcs in a game for example.
I want a model that's as unbiased (in referring to itself as a certain ai model or similar) as possible.

Using the following datasets:

First, Qwen3 1.7b base was trained on the datasets to add basic chat and instruct capabilities, for 20 epochs. This created a very incomplete model.
Then, Qwen3 1.7b's (instruct) LM head was added to the model, and the config from instruct was used. Running the model now still gave broken outputs but attempted to use the actual chat tokens.
Finally, this modified model was finetuned for another 20 epochs on the datasets, after which the model became usable, but still underfitted as of now.

Adding a dataset with more thinking could assist the model with thinking tasks. More training is still needed. Using the smallest dataset size as the step count cap for the datasets would allow for equal use of each dataset instead of the current situation where function calling is used the most.