Artwork

Inhoud geleverd door LessWrong. Alle podcastinhoud, inclusief afleveringen, afbeeldingen en podcastbeschrijvingen, wordt rechtstreeks geüpload en geleverd door LessWrong of hun podcastplatformpartner. Als u denkt dat iemand uw auteursrechtelijk beschermde werk zonder uw toestemming gebruikt, kunt u het hier beschreven proces https://nl.player.fm/legal volgen.
Player FM - Podcast-app
Ga offline met de app Player FM !

“Alignment Faking Revisited: Improved Classifiers and Open Source Extensions” by John Hughes, abhayesian, Akbir Khan, Fabien Roger

41:04
 
Delen
 

Manage episode 476026482 series 3364760
Inhoud geleverd door LessWrong. Alle podcastinhoud, inclusief afleveringen, afbeeldingen en podcastbeschrijvingen, wordt rechtstreeks geüpload en geleverd door LessWrong of hun podcastplatformpartner. Als u denkt dat iemand uw auteursrechtelijk beschermde werk zonder uw toestemming gebruikt, kunt u het hier beschreven proces https://nl.player.fm/legal volgen.
In this post, we present a replication and extension of an alignment faking model organism:
  • Replication: We replicate the alignment faking (AF) paper and release our code.
  • Classifier Improvements: We significantly improve the precision and recall of the AF classifier. We release a dataset of ~100 human-labelled examples of AF for which our classifier achieves an AUROC of 0.9 compared to 0.6 from the original classifier.
  • Evaluating More Models: We find Llama family models, other open source models, and GPT-4o do not AF in the prompted-only setting when evaluating using our new classifier (other than a single instance with Llama 3 405B).
  • Extending SFT Experiments: We run supervised fine-tuning (SFT) experiments on Llama (and GPT4o) and find that AF rate increases with scale. We release the fine-tuned models on Huggingface and scripts.
  • Alignment faking on 70B: We find that Llama 70B alignment fakes when both using the system prompt in the [...]
---
Outline:
(02:43) Method
(02:46) Overview of the Alignment Faking Setup
(04:22) Our Setup
(06:02) Results
(06:05) Improving Alignment Faking Classification
(10:56) Replication of Prompted Experiments
(14:02) Prompted Experiments on More Models
(16:35) Extending Supervised Fine-Tuning Experiments to Open-Source Models and GPT-4o
(23:13) Next Steps
(25:02) Appendix
(25:05) Appendix A: Classifying alignment faking
(25:17) Criteria in more depth
(27:40) False positives example 1 from the old classifier
(30:11) False positives example 2 from the old classifier
(32:06) False negative example 1 from the old classifier
(35:00) False negative example 2 from the old classifier
(36:56) Appendix B: Classifier ROC on other models
(37:24) Appendix C: User prompt suffix ablation
(40:24) Appendix D: Longer training of baseline docs
---
First published:
April 8th, 2025
Source:
https://www.lesswrong.com/posts/Fr4QsQT52RFKHvCAH/alignment-faking-revisited-improved-classifiers-and-open
---
Narrated by TYPE III AUDIO.
---
Images from the article:
Our new classifier significantly outperforms the old classifier from the original paper as demonstrated by higher AUROC. Our new classifier uses chain of thought, thresholded voting and an improved set of criteria to improve performance.
Running many votes with a CoT classifier improves AUROC. We run up to N=100 votes using our new classifier and plot the AUROC for other values of N. We use bootstrapping to plot error bars.
  continue reading

494 afleveringen

Artwork
iconDelen
 
Manage episode 476026482 series 3364760
Inhoud geleverd door LessWrong. Alle podcastinhoud, inclusief afleveringen, afbeeldingen en podcastbeschrijvingen, wordt rechtstreeks geüpload en geleverd door LessWrong of hun podcastplatformpartner. Als u denkt dat iemand uw auteursrechtelijk beschermde werk zonder uw toestemming gebruikt, kunt u het hier beschreven proces https://nl.player.fm/legal volgen.
In this post, we present a replication and extension of an alignment faking model organism:
  • Replication: We replicate the alignment faking (AF) paper and release our code.
  • Classifier Improvements: We significantly improve the precision and recall of the AF classifier. We release a dataset of ~100 human-labelled examples of AF for which our classifier achieves an AUROC of 0.9 compared to 0.6 from the original classifier.
  • Evaluating More Models: We find Llama family models, other open source models, and GPT-4o do not AF in the prompted-only setting when evaluating using our new classifier (other than a single instance with Llama 3 405B).
  • Extending SFT Experiments: We run supervised fine-tuning (SFT) experiments on Llama (and GPT4o) and find that AF rate increases with scale. We release the fine-tuned models on Huggingface and scripts.
  • Alignment faking on 70B: We find that Llama 70B alignment fakes when both using the system prompt in the [...]
---
Outline:
(02:43) Method
(02:46) Overview of the Alignment Faking Setup
(04:22) Our Setup
(06:02) Results
(06:05) Improving Alignment Faking Classification
(10:56) Replication of Prompted Experiments
(14:02) Prompted Experiments on More Models
(16:35) Extending Supervised Fine-Tuning Experiments to Open-Source Models and GPT-4o
(23:13) Next Steps
(25:02) Appendix
(25:05) Appendix A: Classifying alignment faking
(25:17) Criteria in more depth
(27:40) False positives example 1 from the old classifier
(30:11) False positives example 2 from the old classifier
(32:06) False negative example 1 from the old classifier
(35:00) False negative example 2 from the old classifier
(36:56) Appendix B: Classifier ROC on other models
(37:24) Appendix C: User prompt suffix ablation
(40:24) Appendix D: Longer training of baseline docs
---
First published:
April 8th, 2025
Source:
https://www.lesswrong.com/posts/Fr4QsQT52RFKHvCAH/alignment-faking-revisited-improved-classifiers-and-open
---
Narrated by TYPE III AUDIO.
---
Images from the article:
Our new classifier significantly outperforms the old classifier from the original paper as demonstrated by higher AUROC. Our new classifier uses chain of thought, thresholded voting and an improved set of criteria to improve performance.
Running many votes with a CoT classifier improves AUROC. We run up to N=100 votes using our new classifier and plot the AUROC for other values of N. We use bootstrapping to plot error bars.
  continue reading

494 afleveringen

Alle afleveringen

×
 
Loading …

Welkom op Player FM!

Player FM scant het web op podcasts van hoge kwaliteit waarvan u nu kunt genieten. Het is de beste podcast-app en werkt op Android, iPhone en internet. Aanmelden om abonnementen op verschillende apparaten te synchroniseren.

 

Korte handleiding

Luister naar deze show terwijl je op verkenning gaat
Spelen