yufan/user_profile_dataset

👤 User Profile Dataset

Gated chat/SFT data for training MAI Profile models. The repository contains user-profile examples generated by Qwen 3.8 Flash Next and GPT-5.4.

📦 Dataset variants

Teacher	Layer	Configuration	Train	Test
Qwen 3.8 Flash Next	1	User_Profile_L1_qwen38_flash_next	69,000	1,952
Qwen 3.8 Flash Next	2	User_Profile_L2_qwen38_flash_next	73,000	1,327
GPT-5.4	1	User_Profile_L1_gpt54	253,606	5,036
GPT-5.4	2	User_Profile_L2_gpt54	237,245	4,696

The Qwen configurations use deterministic random splits with seed 42. The GPT-5.4 configurations use deterministic 2% user-level test splits, so a user never appears in both train and test within a layer.

🟢 GPT-5.4

🧬 Shared schema

Column	Type	Description
uid	string	Anonymized user identifier
messages	list of {role, content}	User prompt and assistant response
source	string	Constant: GPT54
domain	string	Layer-specific constant: maiprofile-layer1-combined or maiprofile-layer2-combined
think_type	string	Constant: none
language	string	L1 content locale, or L2 market (en/glb)

📊 Published data

Layer	Split	Rows	Users
1	Train	253,606	20,364
1	Test	5,036	418
2	Train	237,245	19,972
2	Test	4,696	410

Layer	Tokenizer	Mean	P50	P90	P95	P99	P99.9	Max
1	Qwen/Qwen3.5-4B	4,002	2,647	9,589	10,399	12,169	13,830	23,589
2	Qwen/Qwen3.5-4B	2,632	2,437	4,656	4,903	5,305	5,905	7,807

🌐 Layer 1 language distribution

Language	Train	Test
en	151,476	3,139
mix	24,857	438
de	18,547	359
ja	13,892	222
fr	12,681	280
pt	9,035	175
es	8,777	166
nl	4,103	124
zh-Hans	3,898	40
zh-Hant	2,131	5
pl	1,628	24
it	1,363	48
sv	1,218	16

💰 GPT-5.4 generation cost

🔹 Layer 1

Stage	Locale	System prompt	Avg input	Avg output	Avg total	Cache rate
l1_delta	en-star	3,435	7,569	493	8,062	44.0%
l1_delta	global	3,435	8,577	543	9,120	38.9%
l1_delta	all	3,435	8,066	518	8,584	41.3%
l1_actual	en-star	601	1,483	327	1,810	0.0%
l1_actual	global	601	1,749	358	2,107	0.0%
l1_actual	all	601	1,614	342	1,956	0.0%
l1_intent	en-star	1,116	1,575	232	1,807	65.0%
l1_intent	global	1,116	1,618	246	1,864	63.3%
l1_intent	all	1,116	1,596	239	1,835	64.1%

🔸 Layer 2

Stage	Locale	System prompt	Avg input	Avg output	Avg total	Cache rate
l2_merge	en-star	2,144	3,943	699	4,642	52.2%
l2_merge	global	2,144	4,062	748	4,810	50.6%
l2_merge	all	2,144	3,993	719	4,712	51.5%
l2_temporal	en-star	1,017	1,955	355	2,310	0.3%
l2_temporal	global	1,017	2,058	375	2,433	0.4%
l2_temporal	all	1,017	2,006	365	2,371	0.4%

🧹 GPT-5.4 data quality

📈 Interest count

Locale	Rows	Mean	P50	P90	P95	P99	P99.9	Max
en-star	140,309	7.3	6	15	17	23	37	87
global	136,918	7.8	7	15	17	25	43	71
all	277,227	7.5	7	15	17	24	40	87

🎯 Interest-count filtering thresholds

Drop if count exceeds	Dropped rows	Drop rate
15	21,807	7.87%
20	5,996	2.16%
25	2,112	0.76%
30	981	0.35%
40 (approximately P99.9)	275	0.10%

✅ Cleaning summary

Result	en-star	en-star %	global	global %	Total	Total %
Rows dropped	7,116	5.072%	7,357	5.373%	14,473	5.221%
Rows kept	133,193	94.928%	129,561	94.627%	262,754	94.779%

🔍 Cleaning rules

#	Rule	en-star	global	Total
R1	Topic source does not match the sources of its evidence signals	5,103	6,094	11,197
R2	Duplicate (Source, Action) signal set, ignoring Date, idx, and intent	1,539	742	2,281
R3	Row has at least 40 interests	113	199	312
R4	Same input has conflicting outputs	266	27	293
R5	Interest or topic has unexpected keys	13	232	245
R6	Non-English/non-Latin interest_name	26	28	54
R7	Duplicate normalized interest_name within a row	28	18	46
R8	Exact duplicate input and output	29	5	34
R9	Empty required text or an interest with no topics	2	21	23
R10	Duplicate value within a topic's source	7	11	18
R11	Duplicate normalized topic name within an interest	6	9	15
R12	Empty/invalid evidence or evidence references an unknown idx	2	6	8
R13	Duplicate idx within a topic's evidence	1	6	7