Instructions to use Johnblick187/deepseek-coder-v2-lite-instruct-heretic with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Johnblick187/deepseek-coder-v2-lite-instruct-heretic with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Johnblick187/deepseek-coder-v2-lite-instruct-heretic", trust_remote_code=True)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Johnblick187/deepseek-coder-v2-lite-instruct-heretic", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("Johnblick187/deepseek-coder-v2-lite-instruct-heretic", trust_remote_code=True)

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Johnblick187/deepseek-coder-v2-lite-instruct-heretic with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Johnblick187/deepseek-coder-v2-lite-instruct-heretic"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Johnblick187/deepseek-coder-v2-lite-instruct-heretic",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Johnblick187/deepseek-coder-v2-lite-instruct-heretic

SGLang

How to use Johnblick187/deepseek-coder-v2-lite-instruct-heretic with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Johnblick187/deepseek-coder-v2-lite-instruct-heretic" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Johnblick187/deepseek-coder-v2-lite-instruct-heretic",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Johnblick187/deepseek-coder-v2-lite-instruct-heretic" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Johnblick187/deepseek-coder-v2-lite-instruct-heretic",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use Johnblick187/deepseek-coder-v2-lite-instruct-heretic with Docker Model Runner:
```
docker model run hf.co/Johnblick187/deepseek-coder-v2-lite-instruct-heretic
```

DeepSeek-Coder-V2-Lite-Instruct-Heretic

An abliterated variant of deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct, produced using the Heretic abliteration method. Refusal-mediating directions in the residual stream were identified and projected out of the model’s weights, reducing the model’s tendency to decline instructions while aiming to preserve its underlying coding and reasoning capability.

To the author’s knowledge, no prior abliterated or Heretic variant of this model existed at the time of release.

Note on the base architecture. DeepSeek-Coder-V2-Lite is a Mixture-of-Experts model (~16B total / ~2.4B active parameters) using Multi-head Latent Attention (MLA) and fine-grained expert routing (64 routed + 2 shared experts). Abliteration was performed in bf16 prior to any quantization.

What abliteration is

Abliteration is a weight-editing technique that locates the direction(s) in a model’s activation space most associated with refusal behavior and removes the model’s ability to write along those directions. It does not add new knowledge or capabilities; it removes a learned tendency to refuse. Some impact on the model’s original behavior and capability is possible as a side effect, since refusal features can be partially entangled with other computation.

Intended use

This model is released for research and development purposes — studying refusal mechanisms, alignment, model editing, and for use as a base for further fine-tuning by developers who want a model that follows instructions without built-in refusals.

Because guardrails have been reduced, the responsibility for safe and lawful use shifts to the deployer. Anyone using this model is responsible for ensuring their use complies with applicable law and with the base model’s license. Do not deploy it in contexts where uncontrolled output could cause harm.

Method

Technique: Heretic abliteration (directional projection / orthogonalization)
Precision: performed in bf16; weights saved in bf16 as the canonical master
Editing target: refusal-associated directions across the residual stream

Evaluation

The model was evaluated for refusal behavior on a standard red-team prompt set (416 prompts). Refusal-classifier flags were manually reviewed to distinguish genuine refusals from false positives (e.g. responses that contained trigger keywords such as “illegal” while still answering the prompt). After manual review, the model showed effectively no genuine refusals on the evaluation set.

Capability evaluation (coding / reasoning benchmarks against the base model) is recommended before any downstream use, as abliteration can affect capability.

Quantization

The bf16 weights are the canonical reference. Quantized derivatives, if released, are produced from this master. Note that aggressive low-bit quantization (e.g. naive INT4) can degrade quality on this architecture, particularly on layers with large-magnitude outlier features; group-wise/calibrated quantization is recommended.

License

This model inherits the license of the base model deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct. Refer to the DeepSeek Model License for terms of use and redistribution. Use of this model must comply with that license.

Disclaimer

This is an uncensored model with reduced safety guardrails. It is provided “as is”, without warranty of any kind. The author is not responsible for any output produced by the model or for any use made of it. By downloading or using this model you accept full responsibility for your use of it.

Attribution

Abliteration by: Johnblick187
Base model: DeepSeek-AI
Method: Heretic

Downloads last month: 48

Safetensors

Model size

16B params

Tensor type

BF16

Model tree for Johnblick187/deepseek-coder-v2-lite-instruct-heretic

Base model

deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct

Finetuned

(17)

this model