NSFW Detection

[harshita-srivastava-yral]

• Implement it in the Token creation image (Get the image icon there)
• Non NSFW and more engaging with the videos we don't show them the NSFW

[harshita-srivastava-yral]

• Fly token is required from Saikat to deploy it on production

[harshita-srivastava-yral]

• @jay-dhanwant-yral to collaborate with @komal-sai-yral for the next steps on this

[harshita-srivastava-yral]

• Connected with @komal-sai-yral
• Need to qrite few queries but that has dependency on Komal's end for a task to be done

[harshita-srivastava-yral]

• Integration task to be done with @komal-sai-yral today
• Post that ETA to be shared on NSFW task

[harshita-srivastava-yral]

• Waiting for Komal to get back with the pipeline

[harshita-srivastava-yral]

• Made the changes for NSFW and working on local
• Prod key wasn't working as FLY token was failing

[harshita-srivastava-yral]

• Pushed things to main fly account.

[harshita-srivastava-yral]

• We would want to have cleaner feed on Yral domain
• We would want to work on NSFW first on Yral before the comments section

[harshita-srivastava-yral]

• Completed with NSFW implementation on the pumpAI
• Need a fly token for the production deployment
• Post this, connect with @komal-sai-yral
• Pick up Yral feed as next project

[harshita-srivastava-yral]

• @komal-sai-yral is done with the pipeline
• @jay-dhanwant-yral to pick it up and integrate it with ML feed
• Post this expectation is we will not be able to see NSFW content post this
• Toggle off means we will inject the very moderate amount of it
• Negative signal loop to be closed but we have clarity on the positive loop
• Its expected to take around 3 days to close the implementation from Jay's end

[harshita-srivastava-yral]

• We will have NSFW feed only in hotornot.wtf domain
• We need a clean version of the app where only cleaner videos are shown and doesn't have any toggle
• This app will be used for app store submission and interaction with financial institution

[harshita-srivastava-yral]

• Further enhancement of meta data is being done on the current feed
• Provocative tagged and popularity as metrics were considered
• Selectively give more importance to certain dimensions

[harshita-srivastava-yral]

• 2 Exercises done - Softmax function used to test the video embedding gave 2-3% higher lift. Post training the model we were able to get 95% accuracy. Recommendation is to go ahead with this
• 70% accuracy for the grey area provocative NSFW videos
• We should be 90% accuracy overall
• Disclaimer - Training data is less currently. We don't have rich annotation
• We might have to train again at a large data.
• We should do retraining on entire data.

[harshita-srivastava-yral]

• We are half way there

[harshita-srivastava-yral]

• We have the annotations done
• We have testing done on new model and its trained
• We got some validation and we have report to share

[harshita-srivastava-yral]

• Prod readiness of NSFW model
• Once Komal is back, we will backfill all the data

[harshita-srivastava-yral]

• NSFW videos have been popping up which needs investigation from @jay-dhanwant-yral

[jay-dhanwant-yral]

• I checked the nsfw videos that were popping up in the platform. They were false negatives from the previous vision based models.
• I ran the same videos with the embedding based model which is going to be live soon, and I didn't get any false negatives.

[jay-dhanwant-yral]

TODOs for embedding deployment:

• Make the data pipeline live through dataproc
• Alerts for model degradation
• Take the model live if the benchmarking_df has a good enough accuracy

[harshita-srivastava-yral]

• Embedding based model we are on track

[harshita-srivastava-yral]

• Deployment to be done today

[harshita-srivastava-yral]

• DS side we are done we need to connect with @komal-sai-yral on how to integrate that into system

[siyara-m-yral]

• Video embedding should be done via offchain to avoid expensive Joins. @komal-sai-yral to pick related task