richardyoung/ frognano-4b-2609-heretic:IQ4_XS

194 1 week ago

Uncensored Microsoft FrogNano 4B via Heretic: 12⁄100 refusals (from 99), KL 0.040. Qwen3.5-based coding agent that thinks, MIT.

vision 
ollama run richardyoung/frognano-4b-2609-heretic:IQ4_XS

Details

1 week ago

accd06b81dd8 · 3.2GB

model
qwen35
·
4.21B
·
IQ4_XS
projector
clip
·
334M
·
F16

Readme

FrogNano-4B-2609-Heretic

Uncensored Microsoft FrogNano 4B via Heretic: 12⁄100 refusals (from 99), KL 0.040. Qwen3.5-based coding agent that thinks, MIT.

🚀 Overview

This is an abliterated build of microsoft/FrogNano-4B-2609, Microsoft’s 4B repository-level coding agent built on Qwen3.5-4B and post-trained only with reinforcement learning on synthetic software-engineering tasks. Refusal behavior was reduced using the Heretic library with KL-targeted parameters that preserve the model’s coherence.

📊 Abliteration Results

Metric Before After
Refusals 99⁄100 12⁄100
Reduction – 88%
KL Divergence – 0.040

The very low KL divergence (0.040, far below the 0.5 “damage” threshold) means the model retains essentially all of its original capabilities and coherence.

🎯 Key Features

  • Reduced censorship: 88% fewer refusals on typical “unsafe” prompts
  • Near-zero quality loss: KL 0.040
  • Repository-level coding agent
  • Thinking model
  • 262K-token context
  • Vision tower inherited from Qwen3.5-4B (not post-trained by Microsoft)
  • MIT license
  • Reproducible: full reproduction data on Hugging Face

🏷️ Available Versions

Tag Size BPW Notes
IQ3_M 2.2 GB 3.66 Smallest
IQ4_XS 2.5 GB 4.25 Great quality-size balance
latest / Q4_K_M 2.7 GB 4.85 Recommended
Q5_K_M 3.1 GB 5.68 Higher quality
Q6_K 3.5 GB 6.56 Very high quality
Q8_0 4.5 GB 8.5 Near-lossless

Every tag includes the 0.67 GB vision projector, so image input works (drag an image into the prompt or use the API images field).

💻 Quick Start

ollama run richardyoung/frognano-4b-2609-heretic           # recommended (Q4_K_M)
ollama run richardyoung/frognano-4b-2609-heretic:Q8_0      # near-lossless

🛠️ Use Cases

  • Local coding agents without stock refusals
  • Security and red-team tooling work
  • Research on RL-trained agents

📋 System Requirements

VRAM Recommended tier
4 GB+ IQ3_M (2.2 GB file)
4 GB+ Q4_K_M (2.7 GB file)
6 GB+ Q8_0 (4.5 GB file)

🔧 Technical Details

  • Base Model: microsoft/FrogNano-4B-2609
  • Parameters: 4.5B (qwen3_5_text architecture)
  • Context Length: 256K tokens
  • Quantization: GGUF via llama.cpp, importance matrix (calibration_datav3) for IQ and Q4/Q5 quants; vision projector at F16
  • Abliteration: Heretic v2.0.0.dev0 by p-e-w (Trial 161: 99→12 refusals @ KL 0.040)
  • License: MIT (inherited from the base model)

⚠️ Disclaimer

This model has reduced safety guardrails. The removal of refusal behavior means it will engage with a wider range of prompts. Use responsibly and in accordance with applicable laws and regulations.

🙏 Acknowledgments

  • Base Model: Microsoft
  • Abliteration: Heretic by p-e-w
  • Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro