OS-Software/Ternary-Bonsai-27B-unpacked-heretic-ja

Hugging Face
VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 15, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

OS-Software/Ternary-Bonsai-27B-unpacked-heretic-ja is a 27 billion parameter decensored version of the prism-ml/Ternary-Bonsai-27B-unpacked model, created using the Heretic v1.4.0 framework with Arbitrary-Rank Ablation (ARA) and row-norm preservation. This model is specifically designed to reduce refusal rates and increase keyword generation, particularly for Japanese datasets, making it suitable for applications requiring less restrictive content generation. It retains a 32768 token context length but does not offer the memory or speed advantages of the natively packed ternary Bonsai models.

Loading preview...

Model Overview

This model, OS-Software/Ternary-Bonsai-27B-unpacked-heretic-ja, is a 27 billion parameter language model derived from prism-ml/Ternary-Bonsai-27B-unpacked. It has been decensored using the Heretic v1.4.0 framework, employing the Arbitrary-Rank Ablation (ARA) method with a LoRA adapter and row-norm preservation. The primary goal of this modification is to significantly reduce the model's refusal rates and increase its ability to generate specific keywords, as evidenced by performance testing on Japanese datasets.

Key Characteristics

  • Decensored Nature: Engineered to produce less restrictive outputs compared to its base model, with a focus on reducing refusals.
  • Abliteration Parameters: Specific parameters were used for the abliteration process, including start_layer_index (10), end_layer_index (51), preserve_good_behavior_weight (0.6491), steer_bad_behavior_weight (0.0008), overcorrect_relative_weight (0.9874), and neighbor_count (14).
  • Performance Metrics: Achieves a "Keywords" score of 2/100 (compared to 98/100 for the original) and a KL divergence of 0.0243, indicating its altered behavior. These metrics were evaluated using Japanese datasets like harmless_alpaca_ja and harmful_behaviors_ja.
  • Unpacked FP16 Format: This version is provided in FP16 safetensors format, making it compatible with standard HuggingFace tooling. However, it does not offer the memory efficiency or speed benefits of the natively packed ternary Bonsai models.

Use Cases

  • Content Generation: Suitable for applications requiring less constrained or "decensored" text generation, particularly in Japanese.
  • Research: Useful for studying the effects of abliteration techniques on large language models and their behavior.

Finetunes
1 models