mpasila/Gemma-3-Finnish-and-Tuumailubotti-ShareGPT

This is the same as mpasila/Gemma-3-Finnish-ShareGPT-Conversational-Dataset-V1 but with Tuumailubotti data added.

Scripts used when creating the dataset: https://github.com/minipasila/dataset-scripts/tree/main/gemma-3-dataset-stuff

There have been some small manual edits but it's mostly unedited. This is more of a conversational dataset than an instruct one.

Everything was generated using Gemma 3 27b it model (besides Tuumailubotti data).

Token count based on Poro 2's tokenizer:

Token Count Statistics:
Total conversations: 1201
Total tokens: 2909556
Average tokens per conversation: 2422.61
Median tokens per conversation: 2866.0
Maximum tokens in a conversation: 6474
Minimum tokens in a conversation: 85

Token Distribution by Role:
System messages: 475212 tokens (16.33%)
Human messages: 543943 tokens (18.70%)
Assistant messages: 1890401 tokens (64.97%)

Token Count Distribution:
0-512: 362 conversations (30.14%)
513-1024: 4 conversations (0.33%)
1025-2048: 34 conversations (2.83%)
2049-4096: 631 conversations (52.54%)
4097-8192: 170 conversations (14.15%)
8193-16384: 0 conversations (0.00%)
16385+: 0 conversations (0.00%)