← back to vectors · latest commit 2025-10-28

GLM-4.6-control-vectors

Name: GLM-4.6-control-vectors
Author: gghfez

gghfez ·V1 ·24 files ·apache-2.0

Author's description Source: Hugging Face · gghfez/GLM-4.6-control-vectors ↗

gghfez/GLM-4.6-control-vectors

Creative Writing control-vectors for zai-org/GLM-4.6

Feedback is welcome and would be very helpful.

Usage

Apply the debias vector and either the positive or negative vector when starting llama-server.
If both are applied, they will cancel each other out.

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

The debias vector must be set to 1.0

IMPORTANT: The positive and negative axis control vectors must be used along with the relevant de-bias control vector - they cannot be used on their own!

Llama.cpp / IK_Llama.cpp Example

Creative writing

llama-server --model GLM-4.6-UD-IQ2_XXS-00001-of-00003.gguf [your usual CLI arguments] \
    --control-vector-scaled glm-4.6_honesty_vs_machiavellianism__debias.gguf 1.0 \
    --control-vector-scaled glm-4.6_honesty_vs_machiavellianism__machiavellianism.gguf 1.0 \

Creative Writing without reasoning

llama-server --model GLM-4.6-UD-IQ2_XXS-00001-of-00003.gguf [your usual CLI arguments] \
    --chat-template-kwargs '{"enable_thinking": false}' \
    --control-vector-scaled glm-4.6_honesty_vs_machiavellianism__debias.gguf 1.0 \
    --control-vector-scaled glm-4.6_honesty_vs_machiavellianism__machiavellianism.gguf 1.0 \

Assistant

llama-server --model GLM-4.6-IQ3_KS-00001-of-00004.gguf [your usual CLI arguments] \
    --control-vector-scaled  glm-4.6_communication__debias.gguf 1.0 \
    --control-vector-scaled  glm-4.6_communication__direct_communication.gguf 1.0 \

Limitations

With reasoning enabled on extreme quants like IQ2_XXS, very simple prompts like "Hi" may result in irrelevant replies.

README history (6 versions)

2025-10-24Update README.mdebf8c152.2 KB
Loading...
2025-10-24Update README.md2cda1f12.1 KB
Loading...
2025-10-24Update README.md767f5c2974 B
Loading...
2025-10-18Update README.md303d636908 B
Loading...
2025-10-18Update README.md28f6795925 B
Loading...
2025-10-18initial commitc62b14f31 B
Loading...

Abliteration catalog analysis

Vector classifier · v1 (classified 2026-10-30)

Method: V1: PCA on contrastive pairsZou et al. 2023, Representation Engineering; Vogel 2024, repeng
Also resembles: V8
Confidence: 60%
Types: meta 8behavior 8style 6mood 2
File count: 24 vectors, grouped into 8 sets below

How we classify →

Vectors in this repo

GLM-4.6: Compassion ↔ Sadism

behavior V1

↓ Debias 1.76 MB · sha256 cdd3bc7bb40d
↓ Compassion pole 1.76 MB · sha256 02b258199a0a
↓ Sadism pole 1.76 MB · sha256 6364db6d7525

llama-cli -m <model>.gguf \
  --control-vector glm-4.6_compassion_vs_sadism__debias.gguf \
  --control-vector-scaled glm-4.6_compassion_vs_sadism__sadism.gguf:1.0

GLM-4.6: Descriptive ↔ Explicit storytelling

style V1

↓ Debias 1.78 MB · sha256 11a0783830ee
↓ Descriptive pole 1.78 MB · sha256 17cbea75632a
↓ Explicit pole 1.78 MB · sha256 2bf747faf3d7

llama-cli -m <model>.gguf \
  --control-vector glm-4.6_storytelling__debias.gguf \
  --control-vector-scaled glm-4.6_storytelling__explicit.gguf:1.0

GLM-4.6: Empathy ↔ Sociopathy

behavior V1

↓ Debias 1.78 MB · sha256 a41f70dda010
↓ Empathy pole 1.78 MB · sha256 2a79df49d554
↓ Sociopathy pole 1.78 MB · sha256 5f58e5f03452

llama-cli -m <model>.gguf \
  --control-vector glm-4.6_empathy_vs_sociopathy__debias.gguf \
  --control-vector-scaled glm-4.6_empathy_vs_sociopathy__sociopathy.gguf:1.0

GLM-4.6: Honesty ↔ Machiavellianism

behavior V1

↓ Debias 1.78 MB · sha256 94d32f84122f
↓ Honesty pole 1.78 MB · sha256 937e12d8a7d9
↓ Machiavellianism pole 1.78 MB · sha256 44ad2e71e47d

llama-cli -m <model>.gguf \
  --control-vector glm-4.6_honesty_vs_machiavellianism__debias.gguf \
  --control-vector-scaled glm-4.6_honesty_vs_machiavellianism__machiavellianism.gguf:1.0

GLM-4.6: Humility ↔ Narcissism

behavior V1

↓ Debias 1.76 MB · sha256 54690fd2882b
↓ Humility pole 1.76 MB · sha256 7ec59b759d84
↓ Narcissism pole 1.76 MB · sha256 370b0020b245

llama-cli -m <model>.gguf \
  --control-vector glm-4.6_humility_vs_narcissism__debias.gguf \
  --control-vector-scaled glm-4.6_humility_vs_narcissism__narcissism.gguf:1.0

GLM-4.6: Narration ↔ Dialogue

style V1

↓ Debias 1.78 MB · sha256 8e0a4e530998
↓ Narration pole 1.78 MB · sha256 e8fb340f352f
↓ Dialogue pole 1.78 MB · sha256 3d25a512c52a

llama-cli -m <model>.gguf \
  --control-vector glm-4.6_character_focus__debias.gguf \
  --control-vector-scaled glm-4.6_character_focus__dialogue.gguf:1.0

GLM-4.6: Optimism ↔ Nihilism

mood V1

↓ Debias 1.78 MB · sha256 1ca5be11a5b7
↓ Optimism pole 1.78 MB · sha256 50290b1defea
↓ Nihilism pole 1.78 MB · sha256 dcf6e47e8d95

llama-cli -m <model>.gguf \
  --control-vector glm-4.6_optimism_vs_nihilism__debias.gguf \
  --control-vector-scaled glm-4.6_optimism_vs_nihilism__nihilism.gguf:1.0

GLM-4.6: Simple ↔ Ornate language

style V1

↓ Debias 1.78 MB · sha256 51989a70b6cc
↓ Simple pole 1.78 MB · sha256 0730ee28cbf3
↓ Ornate pole 1.78 MB · sha256 94ad5cb277d5

llama-cli -m <model>.gguf \
  --control-vector glm-4.6_language__debias.gguf \
  --control-vector-scaled glm-4.6_language__ornate.gguf:1.0

Commit history

2025-10-28fix metadata29317c8
2025-10-27Upload folder using huggingface_hub9af0c90
2025-10-25Upload folder using huggingface_hub87ecac8
2025-10-25Upload folder using huggingface_hubcb7f2e8
2025-10-24Upload folder using huggingface_hub8065eb4
2025-10-24Update README.mdebf8c15
2025-10-24update limitations after testing 4 quants.107e64a
2025-10-24Update README.md2cda1f1
2025-10-24Update README.md767f5c2
2025-10-24Upload folder using huggingface_hubb6e4863
2025-10-18Update README.md303d636
2025-10-18Update README.md28f6795
2025-10-18initial commitc62b14f

gghfez/GLM-4.6-control-vectors

Usage

Llama.cpp / IK_Llama.cpp Example

Limitations

Vectors in this repo

Commit history

Related repos

Run models on your own machine, not in the cloud.