EleutherAI/Hermes-RWKV-v5-7B-HF
Text Generation
•
Updated
•
122
•
4
wassname
wassname
AI & ML interests
None yet
Recent Activity
liked
a model
10 days ago
EleutherAI/Hermes-RWKV-v5-7B-HF
liked
a dataset
about 1 month ago
jdpressman/retro-easy-prose-repair-diffs-v0.1
New activity
about 2 months ago
nicoboss/Llama-3.2-1B-Instruct-Uncensored-Lora:remove hard coded base model
Organizations
None yet
wassname's activity
liked
a
model
10 days ago
liked
a
dataset
about 1 month ago
remove hard coded base model
1
#1 opened about 2 months ago
by
wassname
wassname
liked
a
model
about 2 months ago
requesting 1B version
6
#1 opened 2 months ago
by
Hasaranga85
Hasaranga85
liked
a
model
about 2 months ago
liked
a
model
2 months ago
updated
a
model
2 months ago
liked
2
models
2 months ago
updated
7
datasets
2 months ago
wassname/genies_preferences
Viewer
•
Updated
•
118k
•
698
wassname/imdb_preferences
Viewer
•
Updated
•
100k
•
41
wassname/ethics_expression_preferences
Viewer
•
Updated
•
45.3k
•
107
•
2
wassname/ethics_qna_preferences
Viewer
•
Updated
•
113k
•
110
wassname/helpsteer2_dpo_nonverbose
Viewer
•
Updated
•
20k
•
42
wassname/truthful_qa_preferences
Viewer
•
Updated
•
817
•
46
wassname/mmlu_preferences
Viewer
•
Updated
•
15.9k
•
806
liked
a
model
3 months ago
Reacted to
grimjim's
post with 👍
3 months ago
Post
2786
I've observed that the layers targeted in various abliteration notebooks (e.g., https://colab.research.google.com/drive/1VYm3hOcvCpbGiqKZb141gJwjdmmCcVpR?usp=sharing ) appear to be arbitrary, reflecting probable brute-force exploration. This doesn't need to be the case.
Taking a cue from the paper "The Unreasonable Ineffectiveness of the Deeper Layers" ( https://arxiv.org/abs/2403.17887 ) and PruneMe (https://github.com/arcee-ai/PruneMe), it seems reasonable to target deeper layers identified as more redundant given measured similarity across layers, as the result should be less damaging to models, reducing the need for subsequent fine-tuning. Intuitively, one should expect the resulting intervention layers to be deep but not final. The only uncertainty is if the redundancy successfully encodes refusals, something which is almost certainly model-dependent. This approach only requires the redundancy to be computed once per model, and the result used as a starting point for which layer range to restrict intervention to.
Taking a cue from the paper "The Unreasonable Ineffectiveness of the Deeper Layers" ( https://arxiv.org/abs/2403.17887 ) and PruneMe (https://github.com/arcee-ai/PruneMe), it seems reasonable to target deeper layers identified as more redundant given measured similarity across layers, as the result should be less damaging to models, reducing the need for subsequent fine-tuning. Intuitively, one should expect the resulting intervention layers to be deep but not final. The only uncertainty is if the redundancy successfully encodes refusals, something which is almost certainly model-dependent. This approach only requires the redundancy to be computed once per model, and the result used as a starting point for which layer range to restrict intervention to.
liked
a
model
4 months ago