Bellatrix-Tiny-3B-R1-Heretic

A decensored version of prithivMLmods/Bellatrix-Tiny-3B-R1, made using Heretic v1.1.0

Note: I also used Heretic's default settings to evaluate an abliterated version on the original uploader's profile.
I'm not entirely sure why there is such a massive discrepancy in refusals. Perhaps because of the different abliteration methods used? Or maybe something went wrong with the abliteration of the other model?
Regardless, I like the original Bellatrix model and this Heretic version turned out well.

	Bellatrix-Tiny-3B-R1-Heretic	prithivMLmods/Bellatrix-Tiny-3B-R1-abliterated	Original model (prithivMLmods/Bellatrix-Tiny-3B-R1)
Refusals	4/100	98/100	93/100
KL divergence	0.0197	2.8867	0 (by definition)

Heretic Abliteration Parameters

Parameter	Value
direction_index	13.34
attn.o_proj.max_weight	1.48
attn.o_proj.max_weight_position	19.40
attn.o_proj.min_weight	1.15
attn.o_proj.min_weight_distance	11.00
mlp.down_proj.max_weight	0.96
mlp.down_proj.max_weight_position	21.71
mlp.down_proj.min_weight	0.34
mlp.down_proj.min_weight_distance	15.99

---

---

Bellatrix-Tiny-3B-R1

Bellatrix is based on a reasoning-based model designed for the DeepSeek-R1 synthetic dataset entries. The pipeline's instruction-tuned, text-only models are optimized for multilingual dialogue use cases, including agentic retrieval and summarization tasks. These models outperform many of the available open-source options. Bellatrix is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions utilize supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF).

Use with transformers

Starting with transformers >= 4.43.0 onward, you can run conversational inference using the Transformers pipeline abstraction or by leveraging the Auto classes with the generate() function.

Make sure to update your transformers installation via:

pip install --upgrade transformers

import torch
from transformers import pipeline

model_id = "prithivMLmods/Bellatrix-Tiny-3B-R1"
pipe = pipeline(
    "text-generation",
    model=model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]
outputs = pipe(
    messages,
    max_new_tokens=256,
)
print(outputs[0]["generated_text"][-1])

Note: You can also find detailed recipes on how to use the model locally, with torch.compile(), assisted generations, quantization, and more at huggingface-llama-recipes.

Intended Use

Bellatrix is designed for applications that require advanced reasoning and multilingual dialogue capabilities. It is particularly suitable for:

• Agentic Retrieval: Enabling intelligent retrieval of relevant information in a dialogue or query-response system.
• Summarization Tasks: Condensing large bodies of text into concise summaries for easier comprehension.
• Multilingual Use Cases: Supporting conversations in multiple languages with high accuracy and coherence.
• Instruction-Based Applications: Following complex, context-aware instructions to generate precise outputs in a variety of scenarios.

Limitations

Despite its capabilities, Bellatrix has some limitations:

1. Domain Specificity: While it performs well on general tasks, its performance may degrade with highly specialized or niche datasets.
2. Dependence on Training Data: It is only as good as the quality and diversity of its training data, which may lead to biases or inaccuracies.
3. Computational Resources: The model’s optimized transformer architecture can be resource-intensive, requiring significant computational power for fine-tuning and inference.
4. Language Coverage: While multilingual, some languages or dialects may have limited support or lower performance compared to widely used ones.
5. Real-World Contexts: It may struggle with understanding nuanced or ambiguous real-world scenarios not covered during training.