piotreknow02/Bielik-PL-11B-v3.0-Instruct-heretic

TEXT GENERATIONConcurrent Unit Cost:1Model Size:15BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Jul 8, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The piotreknow02/Bielik-PL-11B-v3.0-Instruct-heretic is a 15 billion parameter instruction-tuned causal language model, a decensored version of the Bielik-PL-11B-v3.0-Instruct model. Developed by SpeakLeash and ACK Cyfronet AGH, it is optimized for the Polish language and other European languages, featuring a context length of 8192 tokens. This model excels at understanding and processing Polish, providing accurate responses for various linguistic tasks, and has been modified to reduce refusals.

Loading preview...

Model Overview

This model, piotreknow02/Bielik-PL-11B-v3.0-Instruct-heretic, is a 15 billion parameter instruction-tuned causal language model. It is a decensored variant of the original speakleash/Bielik-PL-11B-v3.0-Instruct, created using the Heretic v1.4.0 tool. The base model was developed by SpeakLeash and ACK Cyfronet AGH, fine-tuned from Bielik-11B-v3-Base-20250730 with an APT4 tokenizer optimized for Polish.

Key Differentiators

  • Decensored Version: Significantly reduces refusals compared to the original model (0/100 refusals vs. 71/100 for the original), achieved through specific abliteration parameters.
  • Polish Language Optimization: Developed and trained on multilingual text corpora with a strong emphasis on Polish, leveraging Polish large-scale computing infrastructure.
  • Advanced Alignment: Utilizes DPO-Positive and Reinforcement Learning (RL) with GRPO/Dr. GRPO methods for alignment, enhancing analytical capabilities and reducing artificial response length.
  • Reproducible: The model's modifications are reproducible, with details provided in the reproduce directory.

Use Cases

This model is particularly well-suited for applications requiring:

  • Polish Language Processing: Excels in understanding and generating text in Polish, as well as other European languages.
  • Unfiltered Responses: Ideal for use cases where a reduction in model refusals and more direct answers are desired.
  • Linguistic Tasks: Capable of performing a variety of linguistic tasks with high precision, especially in Polish.

Finetunes
2 models