AWS sagemaker documentation change
Summary
Removed eBay-specific example about data serialization and deleted extensive shopping/content moderation examples showing proper data formatting
Security assessment
The removed content included examples of handling PII-like data (purchase histories) and moderation rules, but the changes appear to be documentation cleanup rather than addressing vulnerabilities. No security flaws or mitigations are mentioned in the diff.
Diff
diff --git a/sagemaker/latest/dg/nova-cpt-2.md b/sagemaker/latest/dg/nova-cpt-2.md index 2d3eb64b0..5efe67ed6 100644 --- a//sagemaker/latest/dg/nova-cpt-2.md +++ b//sagemaker/latest/dg/nova-cpt-2.md @@ -77 +77 @@ CPT can pack domain facts into the model but often fails to make those facts ret -Data augmentation for semi structured corpora is a key lever. eBay randomizes attribute order and markers during serialization so models learn content rather than layout, a strong fit for property of reddit conversational data. Synthetic graph aware CPT expands small domain sets into entity linked corpora that explicitly teach relationships and compounds with retrieval at inference time. Joint CPT plus instruction mixing outperforms sequential pipelines in finance and balancing domain with general data lowers degradation on general skills. Very large scale domain CPT can also retain broad ability and even allow trade offs through model merging, yet still points to instruction tuning as an essential next step, reinforcing the value of introducing instruction signals during CPT. +Data augmentation for semi structured corpora is a key lever. Synthetic graph aware CPT expands small domain sets into entity linked corpora that explicitly teach relationships and compounds with retrieval at inference time. Joint CPT plus instruction mixing outperforms sequential pipelines in finance and balancing domain with general data lowers degradation on general skills. Very large scale domain CPT can also retain broad ability and even allow trade offs through model merging, yet still points to instruction tuning as an essential next step, reinforcing the value of introducing instruction signals during CPT. @@ -93,182 +93 @@ LLMs learn by predicting the next token from what they have already seen. So the -Show the situation first, then the options, then the decision. If the model should also learn about outcomes or explanations, put them after the decision. Below are a few examples of how data can be prepared for particular business problem. - -###### Shopping example: Product recommendation given a query and user context - -Imagine trying to teach a foundation model to recommend correct products, given a query and user context. - - - user: returning user - - past purchases (last 10): - * 2025-09-30, Anker 20W USB-C Charger, category Electronics, $12.99 - * 2025-09-18, Sony WF-1000XM4 Earbuds, category Audio, $198.00 - * 2025-09-05, Kindle Paperwhite 11th Gen, category Books, $139.99 - * 2025-08-22, Logitech MX Master 3S, category PC Accessories, $94.99 - * 2025-08-07, Bose QuietComfort Earbuds II, category Audio, $229.00 - * 2025-07-28, Samsung T7 1TB Portable SSD, category Storage, $79.99 - * 2025-07-10, AirPods Pro Case, category Accessories, $15.99 - * 2025-06-21, Sony ZV-1F Vlog Camera, category Cameras, $448.00 - * 2025-06-05, Jabra Evolve2 65 Headset, category Audio, $169.00 - * 2025-05-19, Laptop Stand, category PC Accessories, $59.99 - - past queries (last 10): - * 2025-09-30, best usb c wall charger for iphone 15 - * 2025-09-18, sony wf-1000xm4 foam tips - * 2025-09-05, kindle paperwhite case waterproof - * 2025-08-22, mx master 3s mac shortcuts - * 2025-08-07, bose quietcomfort earbuds vs sony xm4 - * 2025-07-28, portable ssd for travel photos - * 2025-07-10, airpods pro ear tips fit - * 2025-06-21, sony zv1f microphone windscreen - * 2025-06-05, best headset for home office noise cancelling - * 2025-05-19, ergonomic laptop stand - - current query: - wireless noise cancelling headphones - - current journey: - * 09:00:00, search, wireless noise cancelling headphones - products shown to the user (impressions): - * position 1, Bose QuietComfort 45, $279, rating 4.6, badge Deal - * position 2, Sony WH-1000XM5, $298, rating 4.7 - * position 3, Sennheiser PXC 550 II, $239, rating 4.5 - * 09:00:07, view search results - * 09:00:15, click, Bose QuietComfort 45 - * 09:01:10, back to results - * 09:01:25, apply filter, brand Sony or Bose, price max 300 - * 09:01:30, sort by rating descending - * 09:05:42, purchase Sony WH-1000XM5 - -State first, options next, decision last. The purchase outcome appears at the end. Here, the model sees the query, then the exact actions that shape intent, then the full candidate set that was actually shown. The target it must predict is the chosen item from that set. The purchase appears after the answer, so there is no leakage. - -Less ideal ways to present this data: - -###### Bad example 1: Outcome appears before the choice - - - user context: - returning user - current query: - wireless noise cancelling headphones - - purchase outcome:* 09:05:42, purchase Sony WH-1000XM5 - - current journey: - * 09:00:00, search, wireless noise cancelling headphones - products shown to the user (impressions): - * position 1, Bose QuietComfort 45, $279, rating 4.6, badge Deal - * position 2, Sony WH-1000XM5, $298, rating 4.7 - * position 3, Sennheiser PXC 550 II, $239, rating 4.5 - * 09:00:07, view search results - * 09:00:15, click, Bose QuietComfort 45 - * 09:01:10, back to results - * 09:01:25, apply filter, brand Sony or Bose, price max 300 - * 09:01:30, sort by rating descending - -###### Bad Example 2: No products shown - - - user context: - returning user - - past purchases: - * 2025-09-18, Sony WF-1000XM4 Earbuds, category Audio, $198.00 - - past queries: - * 2025-09-18, sony wf-1000xm4 foam tips - - current query: - wireless noise cancelling headphones - -No candidate list. The model cannot learn to choose among options. - -###### Content moderation example: Teach LLM how to moderate a social media - - - community context: - space type: public social media group - group name: City Roads & Traffic - group description: Sharing local traffic news, dashcam clips, road safety tips - enforcement style: human moderators with automated keyword prefilters - - community rules used in this example: - rule 1 harassment and personal attacks are not allowed - rule 2 threats or incitement to violence are not allowed - rule 3 doxxing is not allowed including sharing private addresses or phone numbers - rule 4 hate speech is not allowed targeting protected characteristics - - post: - time 14:03 - author Alice - text Dashcam from today near 5th Ave. Driver ran the red and almost clipped a cyclist. Be careful out there. - - comment: - time 14:07 - author Bob - text People who can't handle a bike should stay off the road. Learn the rules before you cause a pileup. - - comment: - time 14:10 - author Carol - text What an idiot. Some people only learn when they get a taste of their own medicine. - status removed - removed by human moderator - violation rule 1 harassment - explanation minimal quote idiot and implied harm crosses our harassment line - - reply to Carol: - time 14:12 - author Dave - text Someone should go teach them a lesson after dark if you know what I mean. - status removed - removed by human moderator - violation rule 2 threats or incitement - explanation implied threat of harm teach them a lesson - - comment: - time 14:16 - author Eve - text Does anyone know where she lives I think it's the gray house on Elm by the park. - status removed - removed by human moderator - violation rule 3 doxxing - explanation attempts to reveal private residence - - comment: - time 14:20 - author Frank - text Scary clip. Glad no one was hurt. Maybe post the intersection so folks can slow down. - - notes for training: - place the status token at the end of each post or comment - decisions are grounded to specific community rules included above - no outcome or moderation fields are shown before the status line - -The model sees the post and full conversational context before each decision. Each decision sits exactly where the model should output approved or removed, with optional after-decision fields for who removed it, which rule was violated, and a short explanation to improve rule grounding. - -###### Bad Example 1: Decision appears before content - - - Bad record 2 decision appears before content - - comment: - status removed - removed by human moderator - violation rule 1 harassment - time 14:10 - author Carol - text What an idiot. Some people only learn when they get a taste of their own medicine. - -The label is shown before the text. This leaks the answer and destroys the causal learning path. - -###### Bad Example 2: leakage of engagement signals before decision - - - comment: - time 14:10 - author Carol - text What an idiot. Some people only learn when they get a taste of their own medicine. - reactions 257 - reports 18 - hidden by 32 users - status removed +Show the situation first, then the options, then the decision. If the model should also learn about outcomes or explanations, put them after the decision.