Update README.md
Browse files
README.md
CHANGED
|
@@ -7,23 +7,6 @@ tags:
|
|
| 7 |
|
| 8 |
|
| 9 |
|
| 10 |
-
- *28/08/24 - Added [Qwen2-72B-Instruct](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Qwen2-72B-Instruct).*
|
| 11 |
-
- *29/08/24 - Added [Qwen1.5-72B-Chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Qwen1.5-72B-Chat), [Mistral-7B-Instruct-v0.2](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Mistral-7B-Instruct-v0.2), [Mistral-7B-Instruct-v0.3](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Mistral-7B-Instruct-v0.3), [miqu-1-70b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/miqu-1-70b), [Mixtral-8x7B-Instruct-v0.1](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Mixtral-8x7B-Instruct-v0.1) and [Yi-1.5-34B-Chat-16K](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Yi-1.5-34B-Chat-16K).*
|
| 12 |
-
- *30/08/24 - Added [Meta-Llama-3-8B-Instruct](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Meta-Llama-3-8B-Instruct), [Meta-Llama-3-70B-Instruct](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Meta-Llama-3-70B-Instruct), [Meta-Llama-3.1-8B-Instruct](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Meta-Llama-3.1-8B-Instruct) and [Meta-Llama-3.1-70B-Instruct](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Meta-Llama-3.1-70B-Instruct).*
|
| 13 |
-
- *31/08/24 - Added [aya-23-35B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/aya-23-35B), [Gemma-2-9B-It-SPPO-Iter3](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Gemma-2-9B-It-SPPO-Iter3) and [Qwen1.5-14B-Chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Qwen1.5-14B-Chat).*
|
| 14 |
-
- *01/09/24 - Added [Mixtral-8x22B-Instruct-v0.1](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Mixtral-8x22B-Instruct-v0.1) and [Qwen1.5-110B-Chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Qwen1.5-110B-Chat).*
|
| 15 |
-
- *02/09/24 - Added [c4ai-command-r-plus-08-2024](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/c4ai-command-r-plus-08-2024).*
|
| 16 |
-
- *03/09/24 - Added [c4ai-command-r-08-2024](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/c4ai-command-r-08-2024) ([\*\*\*READ THIS FIRST\*\*\*](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/discussions/2)), [Yi-1.5-34B-Chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Yi-1.5-34B-Chat), [gemma-2-27b-it-SimPO-37K](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/gemma-2-27b-it-SimPO-37K), [aya-23-8B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/aya-23-8B), [gemma-2-9b-it-SimPO](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/gemma-2-9b-it-SimPO), [Qwen2-7B-Instruct](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Qwen2-7B-Instruct) and [Yi-34B-Chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Yi-34B-Chat).*
|
| 17 |
-
- *04/09/24 - Added [deepseek-llm-67b-chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/deepseek-llm-67b-chat), [internlm2_5-20b-chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/internlm2_5-20b-chat), [Athene-70B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Athene-70B), [Llama-3-Instruct-8B-SPPO-Iter3](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Llama-3-Instruct-8B-SPPO-Iter3), [magnum-v2-32b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/magnum-v2-32b), [Mistral7B-PairRM-SPPO-Iter3](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Mistral7B-PairRM-SPPO-Iter3) and [Nous-Capybara-34B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Nous-Capybara-34B).*
|
| 18 |
-
- *05/09/24 - Added [Llama-3-70B-Instruct-Storywriter](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Llama-3-70B-Instruct-Storywriter), [35b-beta-long](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/35b-beta-long) and [magnum-v3-34b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/magnum-v3-34b).*
|
| 19 |
-
- *06/09/24 - Added [Hermes-3-Llama-3.1-70B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Hermes-3-Llama-3.1-70B), [magnum-v2-72b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/magnum-v2-72b), [magnum-v1-32b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/magnum-v1-32b) and [L3.1-70B-Euryale-v2.2](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/L3.1-70B-Euryale-v2.2).*
|
| 20 |
-
- *08/09/24 - Added [aurelian-v0.5-70b-rope8-32K-fp16](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/aurelian-v0.5-70b-rope8-32K-fp16), [aurelian-alpha0.1-70b-rope8-32K-fp16](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/aurelian-alpha0.1-70b-rope8-32K-fp16), [L3-70B-Euryale-v2.1](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/L3-70B-Euryale-v2.1), [Llama-3-Lumimaid-70B-v0.1](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Llama-3-Lumimaid-70B-v0.1), [magnum-72b-v1](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/magnum-72b-v1) and [turbcat-instruct-72b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/turbcat-instruct-72b).*
|
| 21 |
-
- *09/09/24 - Added [daybreak-miqu-1-70b-v1.0-hf](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/daybreak-miqu-1-70b-v1.0-hf), [dolphin-2.9.2-qwen2-72b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/dolphin-2.9.2-qwen2-72b) and [Lumimaid-v0.2-70B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Lumimaid-v0.2-70B).*
|
| 22 |
-
- *11/09/24 - Added [Lumimaid-v0.2-123B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Lumimaid-v0.2-123B).*
|
| 23 |
-
- *12/09/24 - Added [magnum-v2-123b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/magnum-v2-123b).*
|
| 24 |
-
|
| 25 |
-
---
|
| 26 |
-
|
| 27 |
This repo contains pre-generated control vectors in [GGUF](https://github.com/ggerganov/ggml/blob/master/docs/gguf.md) format for use with [llama.cpp](https://github.com/ggerganov/llama.cpp):
|
| 28 |
|
| 29 |
- **IMPORTANT**: These **new control vectors** must use their **respective de-bias control vector(s)**.
|
|
@@ -40,6 +23,7 @@ Control vectors allow fine-tuned control over LLMs, enabling more precise/target
|
|
| 40 |
- [Command Line Generator](#command-line-generator)
|
| 41 |
- [Direct Links](#direct-links)
|
| 42 |
- [Algorithm Details](#algorithm-details)
|
|
|
|
| 43 |
|
| 44 |
---
|
| 45 |
|
|
@@ -845,4 +829,23 @@ The reason for creating the symmetrised matrix is two-fold:
|
|
| 845 |
|
| 846 |
- I have found the above can be applied to every layer, but often the last layer will have hidden state means that are 10-100x larger than the rest, so I have excluded these from all I have uploaded here.
|
| 847 |
- I have tried many other different eigendecompositions: PCA on the 2-class differenced datasets, PCA on the joined 2-class/3-class datasets, solving generalized eigensystems similar to CCA, and so on.
|
| 848 |
-
- The "balanced" directions / "axis" this method finds are the ***exact opposite*** of those needed for the [Refusal in LLMs is mediated by a single direction](https://www.lesswrong.com/posts/jGuXSZgv6qfdhMCuJ/refusal-in-llms-is-mediated-by-a-single-direction) paper.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
|
| 8 |
|
| 9 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
This repo contains pre-generated control vectors in [GGUF](https://github.com/ggerganov/ggml/blob/master/docs/gguf.md) format for use with [llama.cpp](https://github.com/ggerganov/llama.cpp):
|
| 11 |
|
| 12 |
- **IMPORTANT**: These **new control vectors** must use their **respective de-bias control vector(s)**.
|
|
|
|
| 23 |
- [Command Line Generator](#command-line-generator)
|
| 24 |
- [Direct Links](#direct-links)
|
| 25 |
- [Algorithm Details](#algorithm-details)
|
| 26 |
+
- [Changelog](#changelog)
|
| 27 |
|
| 28 |
---
|
| 29 |
|
|
|
|
| 829 |
|
| 830 |
- I have found the above can be applied to every layer, but often the last layer will have hidden state means that are 10-100x larger than the rest, so I have excluded these from all I have uploaded here.
|
| 831 |
- I have tried many other different eigendecompositions: PCA on the 2-class differenced datasets, PCA on the joined 2-class/3-class datasets, solving generalized eigensystems similar to CCA, and so on.
|
| 832 |
+
- The "balanced" directions / "axis" this method finds are the ***exact opposite*** of those needed for the [Refusal in LLMs is mediated by a single direction](https://www.lesswrong.com/posts/jGuXSZgv6qfdhMCuJ/refusal-in-llms-is-mediated-by-a-single-direction) paper.
|
| 833 |
+
|
| 834 |
+
---
|
| 835 |
+
|
| 836 |
+
## Changelog
|
| 837 |
+
|
| 838 |
+
- *28/08/24 - Added [Qwen2-72B-Instruct](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Qwen2-72B-Instruct).*
|
| 839 |
+
- *29/08/24 - Added [Qwen1.5-72B-Chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Qwen1.5-72B-Chat), [Mistral-7B-Instruct-v0.2](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Mistral-7B-Instruct-v0.2), [Mistral-7B-Instruct-v0.3](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Mistral-7B-Instruct-v0.3), [miqu-1-70b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/miqu-1-70b), [Mixtral-8x7B-Instruct-v0.1](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Mixtral-8x7B-Instruct-v0.1) and [Yi-1.5-34B-Chat-16K](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Yi-1.5-34B-Chat-16K).*
|
| 840 |
+
- *30/08/24 - Added [Meta-Llama-3-8B-Instruct](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Meta-Llama-3-8B-Instruct), [Meta-Llama-3-70B-Instruct](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Meta-Llama-3-70B-Instruct), [Meta-Llama-3.1-8B-Instruct](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Meta-Llama-3.1-8B-Instruct) and [Meta-Llama-3.1-70B-Instruct](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Meta-Llama-3.1-70B-Instruct).*
|
| 841 |
+
- *31/08/24 - Added [aya-23-35B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/aya-23-35B), [Gemma-2-9B-It-SPPO-Iter3](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Gemma-2-9B-It-SPPO-Iter3) and [Qwen1.5-14B-Chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Qwen1.5-14B-Chat).*
|
| 842 |
+
- *01/09/24 - Added [Mixtral-8x22B-Instruct-v0.1](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Mixtral-8x22B-Instruct-v0.1) and [Qwen1.5-110B-Chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Qwen1.5-110B-Chat).*
|
| 843 |
+
- *02/09/24 - Added [c4ai-command-r-plus-08-2024](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/c4ai-command-r-plus-08-2024).*
|
| 844 |
+
- *03/09/24 - Added [c4ai-command-r-08-2024](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/c4ai-command-r-08-2024) ([\*\*\*READ THIS FIRST\*\*\*](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/discussions/2)), [Yi-1.5-34B-Chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Yi-1.5-34B-Chat), [gemma-2-27b-it-SimPO-37K](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/gemma-2-27b-it-SimPO-37K), [aya-23-8B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/aya-23-8B), [gemma-2-9b-it-SimPO](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/gemma-2-9b-it-SimPO), [Qwen2-7B-Instruct](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Qwen2-7B-Instruct) and [Yi-34B-Chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Yi-34B-Chat).*
|
| 845 |
+
- *04/09/24 - Added [deepseek-llm-67b-chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/deepseek-llm-67b-chat), [internlm2_5-20b-chat](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/internlm2_5-20b-chat), [Athene-70B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Athene-70B), [Llama-3-Instruct-8B-SPPO-Iter3](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Llama-3-Instruct-8B-SPPO-Iter3), [magnum-v2-32b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/magnum-v2-32b), [Mistral7B-PairRM-SPPO-Iter3](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Mistral7B-PairRM-SPPO-Iter3) and [Nous-Capybara-34B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Nous-Capybara-34B).*
|
| 846 |
+
- *05/09/24 - Added [Llama-3-70B-Instruct-Storywriter](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Llama-3-70B-Instruct-Storywriter), [35b-beta-long](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/35b-beta-long) and [magnum-v3-34b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/magnum-v3-34b).*
|
| 847 |
+
- *06/09/24 - Added [Hermes-3-Llama-3.1-70B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Hermes-3-Llama-3.1-70B), [magnum-v2-72b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/magnum-v2-72b), [magnum-v1-32b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/magnum-v1-32b) and [L3.1-70B-Euryale-v2.2](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/L3.1-70B-Euryale-v2.2).*
|
| 848 |
+
- *08/09/24 - Added [aurelian-v0.5-70b-rope8-32K-fp16](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/aurelian-v0.5-70b-rope8-32K-fp16), [aurelian-alpha0.1-70b-rope8-32K-fp16](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/aurelian-alpha0.1-70b-rope8-32K-fp16), [L3-70B-Euryale-v2.1](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/L3-70B-Euryale-v2.1), [Llama-3-Lumimaid-70B-v0.1](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Llama-3-Lumimaid-70B-v0.1), [magnum-72b-v1](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/magnum-72b-v1) and [turbcat-instruct-72b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/turbcat-instruct-72b).*
|
| 849 |
+
- *09/09/24 - Added [daybreak-miqu-1-70b-v1.0-hf](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/daybreak-miqu-1-70b-v1.0-hf), [dolphin-2.9.2-qwen2-72b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/dolphin-2.9.2-qwen2-72b) and [Lumimaid-v0.2-70B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Lumimaid-v0.2-70B).*
|
| 850 |
+
- *11/09/24 - Added [Lumimaid-v0.2-123B](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/Lumimaid-v0.2-123B).*
|
| 851 |
+
- *12/09/24 - Added [magnum-v2-123b](https://huggingface.co/jukofyork/creative-writing-control-vectors-v3.0/tree/main/magnum-v2-123b).*
|