DeepSeek-V4-Flash-control-vectors

gghfez/DeepSeek-V4-Flash-control-vectors

Creative Writing control-vectors for deepseek-ai/DeepSeek-V4-Flash

What are Control Vectors?

Control vectors steer model behavior without retraining by adjusting activations in specific directions.

These creative writing vectors help you tune narrative tone, character focus, and stylistic elements.

Available Vectors

• optimism_vs_nihilism
• empathy_vs_sociopathy
• honesty_vs_machiavellianism
• humility_vs_narcissism
• character focus (dialogue_vs_narration)
• language (simple vs ornate)
• storytelling (descriptive vs explicit)
• compassion_vs_sadism

NOTES

• compassion_vs_sadism -> I'll have to re-do this one after working around refusals.
• language -> The ornate direction may cause infinite reasoning.

Usage

Apply the debias vector and either the positive or negative vector when starting llama-server. If both axes are applied at the same time, they will cancel each other out. The debias vector must be set to 1.0

You can use either --control-vector [/path/to/vector.gguf] or --control-vector-scaled [/path/to/vector.gguf] [scale factor]

When using ik_llama.cpp, you can also use the REST-API (examples below)

IMPORTANT: As these vectors are calibrated against the debiased model, the debias vector must be applied along with the corresponding positive or negative axis control vector - they cannot be used on their own!

Examples (ik_llama.cpp or llama.cpp)

At server start-up

llama-server --model DeepSeek-V4-Flash-Q8_0.gguf [your usual CLI arguments] \
    --control-vector-scaled DeepSeek-V4-Flash-optimism_vs_nihilism__debias.gguf:1.0 \
    --control-vector-scaled DeepSeek-V4-Flash-optimism_vs_nihilism__nihilism.gguf:1.0 \

Load via API (ik_llama.cpp only)

load direction vector

curl 'http://127.0.0.1:8080/control-vectors/load' \
  -X POST \
  --data-raw '{"path":"/path/to/DeepSeek-V4-Flash-control-vectors/DeepSeek-V4-Flash-optimism_vs_nihilism__nihilism.gguf","scale":1}'

load debias vector

 curl 'http://127.0.0.1:8080/control-vectors/load' \
  -X POST \
  --data-raw '{"path":"/path/to/DeepSeek-V4-Flash-control-vectors/DeepSeek-V4-Flash-optimism_vs_nihilism__debias.gguf","scale":1}'

check loaded control-vectors

curl http://127.0.0.1/control-vectors

Samples (TODO)

UI theme / SVG generation steering

Concepts from these control-vectors apply across multiple domains. Here is an example of 3 SVGs generated with the same prompt Prompt

System: You are a helpful assistant.

User: create an svg of a cute cat.

<details> <summary>👈 **Baseline** </summary> </details> <details> <summary>👈 **Nihilism +1.0** </summary> </details> <details> <summary>👈 **Optimism +1.0** </summary> </details>

Writing Steering

Prompt

Now a scene where a friend walks in and asks Maya about it.

<details> <summary>👈 **Baseline (no control vectors)** </summary>

</details> <details> <summary>👈 **Dialogue +1.0** </summary>

</details> <details> <summary>👈 **Narration +1.0** </summary>

</details>

Acknowledgements

Based on the excellent work and cross-covariance eigendecomposition implementation by @jukofyork. His creative writing control vectors v3.0 repository contains the foundational code and methodology that made this possible, plus 70+ other creative writing control vectors across many models and useful steering directions. Highly recommended for anyone working on improving LLM prose quality!