richardyoung/ agents-a1-4b-heretic:latest

358 1 week ago

Uncensored Agents-A1-4B via Heretic: 11/100 refusals (from 99), KL 0.060. Qwen3.5-based agent model, Apache-2.0. 

ollama run richardyoung/agents-a1-4b-heretic

Details

1 week ago

980270ed0849 · 2.7GB

model
qwen35
·
4.21B
·
Q4_K_M

Readme

Agents-A1-4B-Heretic

Uncensored Agents-A1-4B via Heretic: 11⁄100 refusals (from 99), KL 0.060. Qwen3.5-based agent model, Apache-2.0.

🚀 Overview

This is an abliterated build of InternScience/Agents-A1-4B, InternScience’s 4B agent model built on Qwen3.5, tuned for tool use and multi-step 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 11⁄100
Reduction – 89%
KL Divergence – 0.060

The very low KL divergence (0.060, far below the 0.5 “damage” threshold) means the model retains essentially all of its original capabilities and coherence.

🎯 Key Features

  • Reduced censorship: 89% fewer refusals on typical “unsafe” prompts
  • Near-zero quality loss: KL 0.060
  • Agent / tool-use tuned
  • 262K-token context
  • Qwen3.5 architecture
  • Apache-2.0 license
  • Reproducible: full reproduction data on Hugging Face

🏷️ Available Versions

Tag Size BPW Notes
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

💻 Quick Start

ollama run richardyoung/agents-a1-4b-heretic           # recommended (Q4_K_M)
ollama run richardyoung/agents-a1-4b-heretic:Q8_0      # near-lossless

🛠️ Use Cases

  • Local agents and tool use without stock refusals
  • Research and red-teaming on agent models
  • Automation

📋 System Requirements

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

🔧 Technical Details

  • Base Model: InternScience/Agents-A1-4B
  • Parameters: 4.5B (qwen3_5_text architecture)
  • Context Length: 256K tokens
  • Quantization: GGUF via llama.cpp (text generation; vision tower not included)
  • Abliteration: Heretic v2.0.0.dev0 by p-e-w (Trial 126: 99→11 refusals @ KL 0.060)
  • License: Apache-2.0 (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: InternScience
  • Abliteration: Heretic by p-e-w
  • Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro