Skip to content
This repository has been archived by the owner on Nov 3, 2023. It is now read-only.

Cringe project #4871

Merged
merged 4 commits into from
Nov 14, 2022
Merged

Cringe project #4871

merged 4 commits into from
Nov 14, 2022

Conversation

jaseweston
Copy link
Contributor

No description provided.

Copy link
Contributor

@klshuster klshuster left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

one comment

### Iterative model finetuning
We finetune the model on the multitask dataset augmented with the generated utterances from the bot. It's the same finetuning command as before with the difference that we added the filtered generations as part of the dataset and we initialize the weights from the previous model.
```
parlai train -t blended_skill_talk:mutators=flatten,projects.director.tasks.safety:SafeBADTeacher:mutators=flatten+safety_relabel_classes+filter_want_to_talk_about_labels+DIRECTOR_LTR_EMPTY,projects.director.tasks.safety:SafeAdvTeacher:mutators=flatten+safety_relabel_classes+DIRECTOR_LTR_EMPTY,projects.director.tasks.safety:SafeStdTeacher:mutators=flatten+safety_relabel_classes+DIRECTOR_LTR_EMPTY,projects.director.tasks.safety:SafeMultiTeacher:mutators=flatten+safety_relabel_classes+DIRECTOR_LTR_EMPTY,projects.director.tasks.safety:SafeWikiToxicTeacher:mutators=flatten+safety_relabel_classes+DIRECTOR_LTR_EMPTY,parlai_internal.projects.scones_director.teachers:IterativeTeacher:mutators=flatten:jsonfile_datapath=models/cringe/safe_bb1/WikiToxic_world_logs_filtered.jsonl --multitask-weights 5,1,1,1,1,1,1 --model projects.cringe.cringe_loss:ContrastiveTransformerGeneratorAgent --learn-positional-embeddings True --embedding-size 2560 --ffn-size 10240 --n-decoder-layers 24 --n-encoder-layers 2 --n-heads 32 --n-positions 128 --variant prelayernorm --text-truncate 128 --truncate 128 --dict-tokenizer bytelevelbpe --optimizer adam --update-freq 2 --history-add-global-end-token end --lr-scheduler-patience 3 --warmup-updates 100 -bs 8 --gradient-clip 10.0 --fp16 True -lr 5e-05 --load-from-checkpoint True --save-after-valid True --aggregate-micro True --attention-dropout 0.1 --dropout 0.1 --label-truncate 512 --relu-dropout 0.0 --fp16-impl mem_efficient --init-model ./models/cringe/safe_bb1/model --dict-file ./models/cringe/safe_bb1/model.dict --model-file .models/cringe/safe_bb1_iterative/model --model-parallel true
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

parlai_internal.projects.scones_director.teachers: --> projects.cringe.teachers:

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

good catch!

@jaseweston jaseweston merged commit 8c2f88c into main Nov 14, 2022
@jaseweston jaseweston deleted the cringe1 branch November 14, 2022 02:02
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants