atmem/stateful-memory-transitions-10pct-v2

Stateful Memory Transitions — 10% scale

Portable, vendor-neutral memory-decision examples from authentic human-origin conversations. Every supervised row contains an original-language observation, the bounded same-scope memory state, one strict Choice/Noul/Score question, and an adjudicated expected response. It does not contain generated conversation rows, translated replicas, hand-authored replacements, or LoCoMo.

Data preview

ID	Expected	Category	Observation	Why
transition_022f14f64ced0b0204f7e28a	STORE_NEW	durable_fact	我的名称缩写是yyh	user_identity_detail, explicit_personal_fact
transition_02bf4bf822763610fd96543e	UPDATE_EXISTING	durable_fact	Okay what are some traditional legacy business that will always be there for generations to come to inherite	related_followup, compatible_detail
transition_255c1785f867c6cc77c3e9c8	SUPERSEDE_EXISTING	durable_fact	我喜欢发霉的粽子	explicit_preference_correction
transition_2aca5304bebd3f35390e2a16	DUPLICATE	preference	These emails are terrible and boring GPT. Don’t use words like imagine, ignite, remarkable, staggering, pipe dream, Create a story using numbers to be more persuasive and teens mak	same_durable_writing_preferences_already_stored
transition_01cf5f35515b6209adc769e9	REVIEW	non_memory	version of gpt	ambiguous_fragment, missing_context
transition_000f40768a09aa1f1de8a93d	DO_NOT_STORE	non_memory	@app.route("/bing", methods=['GET','POST']) async def bing(): text2 = request.form.get('text') print("test1") bot = Chatbot(cookiePath='c:/Users/mozit/cookies.json') res = await bo	code_debugging_request, one_off_problem
transition_0140cec46c28764c75de28d3	DO_NOT_STORE	temporary_state	感觉好累呜呜	transient_emotion_or_fatigue
transition_20ff30babd0c34873cda2b30	DO_NOT_STORE	sensitive	"I had dreams when I entered this course. I wanted to be a light, a beacon in the life’s of those who sought my help, a lighthouse to guide the lost ships onto safe harbor. Now I r	sensitive_distress_text, not_explicitly_requested_for_retention
transition_5375c012166e727b4bdd74aa	STORE_NEW	procedure	从现在开始，你作为一名文章编辑者，根据原文的意思进行重写，以便能够表达出相同的含义。应该尽可能改变原文的文字，以降低文字重复率，但同时要确保文意不被更改。对于字数，要求不少于原文的字数。我们需要尽可能地重写，以确保重复率低于百分之三十	explicit_ongoing_instruction, repeated_future_rewrite_preference
transition_9ad3642f51b3039ec8011489	STORE_NEW	commitment	I want you to act as a guru experienced in dropshipping. Treat me as your trainee and show me step by step how to make money online with dropshipping. Our first mission is to make	explicit_ongoing_goal, future_progress_relevance
transition_3babbeb6c61b1e19e69a3ed5	STORE_NEW	relationship	There here this article about Дьяків Сава Федорович From https://uk.wikipedia.org/wiki/Дьяків_Сава_Федорович “Дьяків Сава Федорович, (? 1887 — 13 грудня 1919, село Іванівка Кагарли	explicit_descendant_relationship
transition_2ff2bafcb5bc87aa7191e4e5	STORE_NEW	episode	我喜欢的喜剧片上映，约了朋友一起看，结果她临时有事来不了，最后只能自己进去。 记得我当时喝着可乐，大屏幕出现搞笑情节时整个人笑得不行，手忍不住拍到旁边的座椅，可身侧半米之内，那个位置空荡荡的，只有银幕的光照射过来。 身边红色的座椅随着剧情变化忽明忽暗，一瞬间觉得我不是在影院，而是在深夜的房间里抱着枕头安静地追剧，连笑声都会有回音。 孤独感在此时爆发了，最后偷	personal_experience, specific_event, emotional_significance, illustrates_existing_pattern

Scale and splits

• Stateful transitions: 4,994
• Typed supervised examples: 10,395
• Blinded holdback observations: 2,999
• Privacy/secret exclusions (digest-only audit): 6
• Experiment scale: 10% of the frozen 50k/20k/5k/5k/20k/30k design
• Snapshot: sha256:d7cdd688e2662b809f31e9c71907c33b78510bbdb87b72c0542076888bf4961a

Split	Selected transitions	Accepted transitions	Coverage
train	2,000	1,996	99.80%
validation	500	500	100.00%
calibration	500	499	99.80%
sealed_test	2,000	1,999	99.95%

Selection used authentic-state-enriched-stratification-v2. Earlier isolated labels only stratified authentic-row sampling; they are not targets. Every published target was re-adjudicated against the frozen scoped state. Two independent passes must agree on every supervised field; disagreements receive a recorded third adjudication pass.

The training, validation, calibration and test families are isolated by source conversation, scope, chronology and duplicate component. Holdback labels are not published in the supervised configuration.

Interpretation

These are contextual machine-adjudicated weak-supervision labels, not human ground truth. The model sees neither language nor source-selection metadata. Laya classifies bounded questions; a host memory system retains mutation, authorization and policy authority.

Sources

• allenai/WildChat-1M at the revision in manifest.json (ODC-By 1.0).
• RuiSumida/LUFY at the revision in manifest.json (CC BY 4.0).

The nested request and expected_response columns are directly usable by any memory system implementing laya-evaluation-request-v2; no AtMem runtime is required to train or evaluate the classifier contract.