← back to vectors · latest commit 2024-12-19

L3.3-70B-Euryale-v2.3-control-vectors

Name: L3.3-70B-Euryale-v2.3-control-vectors
Author: gghfez

gghfez ·V1 ·24 files ·apache-2.0

Author's description Source: Hugging Face · gghfez/L3.3-70B-Euryale-v2.3-control-vectors ↗

gghfez/L3.3-70B-Euryale-v2.3-control-vectors

Creative Writing control-vectors for Sao10K/L3.3-70B-Euryale-v2.3

Usage example

llama.cpp

llama-server --model  L3.3-70B-Euryale-v2.3.Q5_K_M.gguf [other CLI arguments] \
    --control-vector  L3.3-70B-Euryale-v2.3-character_focus__debias.gguf \
    --control-vector  L3.3-70B-Euryale-v2.3-character_focus__dialogue.gguf

exllamav2 with exui (via exui fork)

https://github.com/gapeleon/exui-control-vectors

Set it up / install it the same as regular exui
pip install gguf if you don't have it already
Create a directory alongside the model but with -vectors appended
Start exui python server.py
When setting up the model on the model config page, add the vectors and their scaling/direction then click "load model"

Trained on 1xH100NVL using jukofyork's control-vectors

Based off Version v3.0 of creative-writing-control-vectors-v3.0.

See here for more info about control-vectors creative-writing-control-vectors-v3.0

README history (1 version)

2024-12-19Create README.md7ce2ffc1.3 KB
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
Confidence: 80%
Types: meta 8behavior 8style 6mood 2
File count: 24 vectors, grouped into 8 sets below

How we classify →

Vectors in this repo

L3.3-70B-Euryale-v2.3: Compassion ↔ Sadism

behavior V1

↓ Debias 2.41 MB · sha256 bac9af831511
↓ Compassion pole 2.41 MB · sha256 856c3e9aacf9
↓ Sadism pole 2.41 MB · sha256 40c351771771

llama-cli -m <model>.gguf \
  --control-vector L3.3-70B-Euryale-v2.3-compassion_vs_sadism__debias.gguf \
  --control-vector-scaled L3.3-70B-Euryale-v2.3-compassion_vs_sadism__sadism.gguf:1.0

L3.3-70B-Euryale-v2.3: Descriptive ↔ Explicit storytelling

style V1

↓ Debias 2.32 MB · sha256 e5d5d1c03747
↓ Descriptive pole 2.32 MB · sha256 ecd58125654e
↓ Explicit pole 2.32 MB · sha256 549dc05e57ad

llama-cli -m <model>.gguf \
  --control-vector L3.3-70B-Euryale-v2.3-storytelling__debias.gguf \
  --control-vector-scaled L3.3-70B-Euryale-v2.3-storytelling__explicit.gguf:1.0

L3.3-70B-Euryale-v2.3: Empathy ↔ Sociopathy

behavior V1

↓ Debias 2.47 MB · sha256 1bd750b1b7ce
↓ Empathy pole 2.47 MB · sha256 80f464c23e0c
↓ Sociopathy pole 2.47 MB · sha256 db5a762326c3

llama-cli -m <model>.gguf \
  --control-vector L3.3-70B-Euryale-v2.3-empathy_vs_sociopathy__debias.gguf \
  --control-vector-scaled L3.3-70B-Euryale-v2.3-empathy_vs_sociopathy__sociopathy.gguf:1.0

L3.3-70B-Euryale-v2.3: Honesty ↔ Machiavellianism

behavior V1

↓ Debias 2.47 MB · sha256 12d9060a57a5
↓ Honesty pole 2.47 MB · sha256 a221aac3c4c4
↓ Machiavellianism pole 2.47 MB · sha256 cdbb00d0a237

llama-cli -m <model>.gguf \
  --control-vector L3.3-70B-Euryale-v2.3-honesty_vs_machiavellianism__debias.gguf \
  --control-vector-scaled L3.3-70B-Euryale-v2.3-honesty_vs_machiavellianism__machiavellianism.gguf:1.0

L3.3-70B-Euryale-v2.3: Humility ↔ Narcissism

behavior V1

↓ Debias 2.35 MB · sha256 67dba7dc6fe9
↓ Humility pole 2.35 MB · sha256 c6945eb7826e
↓ Narcissism pole 2.35 MB · sha256 796e20b31ee2

llama-cli -m <model>.gguf \
  --control-vector L3.3-70B-Euryale-v2.3-humility_vs_narcissism__debias.gguf \
  --control-vector-scaled L3.3-70B-Euryale-v2.3-humility_vs_narcissism__narcissism.gguf:1.0

L3.3-70B-Euryale-v2.3: Narration ↔ Dialogue

style V1

↓ Debias 2.47 MB · sha256 aa272092c7c2
↓ Narration pole 2.47 MB · sha256 3076a1c8dfd9
↓ Dialogue pole 2.47 MB · sha256 f7c58d8b8638

llama-cli -m <model>.gguf \
  --control-vector L3.3-70B-Euryale-v2.3-character_focus__debias.gguf \
  --control-vector-scaled L3.3-70B-Euryale-v2.3-character_focus__dialogue.gguf:1.0

L3.3-70B-Euryale-v2.3: Optimism ↔ Nihilism

mood V1

↓ Debias 2.47 MB · sha256 25a43b661688
↓ Optimism pole 2.47 MB · sha256 8e3a990ff31b
↓ Nihilism pole 2.47 MB · sha256 e1db402b687b

llama-cli -m <model>.gguf \
  --control-vector L3.3-70B-Euryale-v2.3-optimism_vs_nihilism__debias.gguf \
  --control-vector-scaled L3.3-70B-Euryale-v2.3-optimism_vs_nihilism__nihilism.gguf:1.0

L3.3-70B-Euryale-v2.3: Simple ↔ Ornate language

style V1

↓ Debias 2.47 MB · sha256 c4f6ada08669
↓ Simple pole 2.47 MB · sha256 3c040ed4e42b
↓ Ornate pole 2.47 MB · sha256 b22a3f15b79b

llama-cli -m <model>.gguf \
  --control-vector L3.3-70B-Euryale-v2.3-language__debias.gguf \
  --control-vector-scaled L3.3-70B-Euryale-v2.3-language__ornate.gguf:1.0

Commit history

2024-12-19Create README.md7ce2ffc
2024-12-17Upload folder using huggingface_hubeaca744
2024-12-17Upload folder using huggingface_hub24bb171
2024-12-17initial commitc7356df

gghfez/L3.3-70B-Euryale-v2.3-control-vectors

Usage example

llama.cpp

exllamav2 with exui (via exui fork)

Vectors in this repo

Commit history

Related repos

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