richardyoung/ deepseek-coder-33b-heretic:latest

917 8 months ago

State-of-the-art coding AI trained on 2T tokens with project-level understanding and no content restrictions for unrestricted code generation. 

ollama run richardyoung/deepseek-coder-33b-heretic

Details

8 months ago

9483261a48e3 · 67GB ·

model
llama
·
33.3B
·
F16
system
# DeepSeek-Coder-33B-Instruct Heretic - Uncensored Edition ## Overview This is an **uncensored versi
params
{ "stop": [ "<|im_start|>", "<|im_end|>" ], "temperature": 0.7, "top
template
{{ if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}{{ if .Prompt }}<|im_start|>user

Readme

DeepSeek-Coder-33B-Heretic

An abliterated, code-focused build of DeepSeek-Coder-33B-Instruct with reduced refusals for unobstructed programming assistance.

🚀 Overview

This is a Heretic-abliterated build of deepseek-ai/deepseek-coder-33b-instruct, a 33B-parameter instruction-tuned coding model trained on a large corpus of source code and natural language. Refusal directions were suppressed using the Heretic library with conservative parameters, lowering refusals while preserving the model’s code-generation quality. The result is a capable coding assistant that engages with a wider range of programming prompts.

📊 Abliteration Results

Metric Before After
Refusals TBD TBD
Reduction – TBD
KL Divergence – TBD

Refusal metrics pending re-measurement.

🎯 Key Features

  • Reduced refusals, engages with a far wider range of coding and technical prompts than the stock instruct model.
  • Strong code generation, inherits DeepSeek-Coder’s proficiency across many programming languages (Python, JavaScript/TypeScript, Go, Rust, Java, C/C++, and more).
  • Fill-in-the-middle (FIM), supports infilling and project-level code completion, in addition to instruction-style chat.
  • GGUF quantization, distributed as a quantized GGUF for efficient local inference via Ollama / llama.cpp.

🏷️ Available Versions

Tag Size Context Notes
latest 67GB 16K High-precision GGUF; maximum fidelity to the base model.

Only the latest tag is currently published. Additional quants can be added on request, the bits-per-weight (BPW) guide below indicates the quality/size trade-off.

BPW reference guide:

Quant BPW Trade-off
IQ3_M 3.66 Smallest; lowest fidelity
IQ4_XS 4.25 Compact; good quality
Q4_K_M 4.85 Balanced quality/size
Q5_K_M 5.68 Higher quality
Q6_K 6.56 Near-lossless
Q8_0 8.50 Highest precision

💻 Quick Start

ollama run richardyoung/deepseek-coder-33b-heretic

🛠️ Use Cases

  • Code generation, completion, and fill-in-the-middle infilling.
  • Refactoring, debugging, and code explanation across many languages.
  • Writing tests, scripts, and project scaffolding.
  • Technical Q&A and programming assistance without stock refusal behavior.

📋 System Requirements

Configuration Approx. VRAM / RAM Performance
latest (67GB GGUF), full GPU ~72GB+ VRAM Best on a single 80GB GPU (e.g. A100/H100) or multi-GPU; fast
latest, partial offload 48GB GPU + system RAM Usable with CPU offload; slower
latest, CPU-only 80GB+ system RAM Functional but slow
Smaller quant (e.g. Q4_K_M, ~20GB) ~24GB VRAM Fits a single 24GB GPU; add on request

🔧 Technical Details

  • Base Model: deepseek-ai/deepseek-coder-33b-instruct
  • Parameters: 33B
  • Context Length: 16,384 tokens (16K)
  • Quantization: GGUF (latest tag, 67GB, high-precision)
  • Abliteration: Heretic (p-e-w), conservative weight modification

⚠️ Disclaimer

This model has reduced safety guardrails and will engage with a wider range of prompts. Use responsibly and in accordance with applicable laws.

🙏 Acknowledgments

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

Built & maintained by Richard Young · DeepNeuro