ya9919/fin-laya-decision-qa

Fin-Laya Bilingual Decision QA

中文 | English

Chinese and English training material for Fin-Laya decision models, part of the 1.4B Persona Social Model project. The dataset contains 4,172,421 observed survey question–answer pairs in each language, covering 1,211 year-specific question/task versions. English records are translations of the same Chinese observations: 8,344,842 language rows represent 4,172,421 underlying QA pairs.

Dataset statistics

Metric	Value	Priority
Number of Samples	2,925,263 train examples per language; 5,850,526 bilingual train rows. All splits: 4,172,421 underlying QA pairs / 8,344,842 language rows.	Core
Number of Tokens	Train: 662,633,773 Chinese / 707,123,625 English input tokens with the actual Fin-Laya multilingual tokenizer and training format. Bilingual train total: 1,369,757,398.	Core
Number of Unique Personas	142,647 unique persona groups across all splits; 99,920 in train. Multiple questions and bilingual copies count once.	Core
Number of Attributes	27 observed input dimensions, covering 35 survey field names after excluding administrative context.	Recommended
Data Sources	CHFS, CGSS, CFPS and CLDS; 18 survey waves, 2010–2022.	Recommended
Number of Characters	Auxiliary readable-text volume across all splits: 889,964,865 Chinese / 3,177,697,730 English Unicode characters.	Auxiliary

Persona groups use pseudonymous survey group keys. CHFS groups linked households, CGSS groups respondents, and CFPS/CLDS keep equal source-ID codes together across waves as a conservative split rule. The CFPS/CLDS files do not provide a verified longitudinal identity bridge, so this operational count does not establish 142,647 distinct real people. The separate CFPS/CLDS profile files contain 145,994 household-wave records.

Attribute counts merge equivalent names and coarse/fine representations of age, education, occupation and geographic location. Available dimensions vary by survey and question; target-related fields remain excluded. The detailed mapping is in attribute inventory.

Token totals count non-padding model inputs, including the formatted question, option IDs/descriptions, serialized profile and structural special tokens; answer targets and metadata are excluded. The main count applies the training limits (max_len=512, head_max_len=256, up to 48 tokens per option). It measures one pass through the available data, with no epoch multiplier.

Split	Chinese input tokens	English input tokens
train	662,633,773	707,123,625
dev	93,217,820	99,458,526
calibration	93,481,949	99,745,715
test	96,009,678	102,493,034
All splits	945,343,220	1,008,820,900

Before truncation, train contains 665,625,280 Chinese / 712,805,827 English input tokens. The English-only Fin-Laya tokenizer gives 667,985,366 encoded train tokens; token counts depend on the tokenizer. Fingerprints, split/source totals and truncation statistics are in dataset statistics and tokenizer fingerprints.

Sources

• Southwestern University of Finance and Economics (SWUFE): China Household Finance Survey (CHFS), waves 2011, 2013, 2015, 2017 and 2019; 87 question versions.
• Renmin University of China (RUC): Chinese General Social Survey (CGSS), wave 2021; 32 question versions.

CHFS and CGSS targets are original respondent answers, including financial-knowledge answers. CFPS and CLDS add observed categories from the supplied harmonized household records, with constructed choice-question templates. Records are not linked across surveys.

• Peking University (PKU): China Family Panel Studies (CFPS), waves 2010, 2012, 2014, 2016, 2018, 2020 and 2022; 1,126,227 QA records per language, including financial-state and occupation/education supervision.
• Sun Yat-sen University (SYSU): China Labor-force Dynamic Survey (CLDS), waves 2011, 2012, 2014, 2016 and 2018; 591,042 QA records per language, including financial-state and occupation/education supervision.

From survey records to decision-model training data

1. Register the questions. Match each survey wave and source column to the original question wording, option codes, and option descriptions. Retain separate question IDs for different waves.
2. Build the respondent state. Join CHFS household, individual and geographic tables by household identifiers, selecting the marked respondent record. Read CGSS respondent profiles directly. Use available age, sex/gender, education, marital status, province, household size and urban/rural residence; CHFS demand-deposit balances are included only where allowed by the target-specific exclusions. Missing inputs remain missing or explicitly marked as not provided.
3. Retain eligible observed answers. Map raw answer codes to the registered option IDs. Exclude missing, skipped, refused or undefined codes rather than inventing answers. An explicit “do not know” option remains a valid response where the questionnaire defines it. Split CGSS multiple-response investment items into individual choice indicators only when the parent item was answered; an unselected option is not inferred from a missing parent response.
4. Create a typed decision example. Each eligible response produces state, questions, expected and meta. The prompt asks for the respondent’s likely answer while retaining their financial-knowledge level. The original answer becomes the training target; target answers and provenance are not inference inputs. Omit target-related input fields according to the recorded exclusions.
5. Freeze group-based splits. Keep linked CHFS households across waves together and CGSS respondents together. Replace original identifiers with pseudonymous group hashes and assign train/dev/calibration/test using the fixed hash rule. All questions for a group, and both language versions, stay in the same split.

The training material supports SFT and native Laya RLCD. The original 21-question subset is listed in metadata/initial_question_ids.json. The frozen CHFS/CGSS component has 119 question versions; CFPS/CLDS add 123 year-specific versions of 13 financial-state tasks. English text must be encoded separately from Chinese text.

CFPS and CLDS preparation

The supplied files contain 148,537 household-wave rows and 44 aligned columns. The 2012/2014 CFPS files include 2,543 empty placeholder rows with missing household IDs; 2,541 are exact repetitions. Removing all placeholders leaves 145,994 household-wave records, which produce 1,461,778 additional QA records. The newer 44-column extract extends education and occupation information; its financial values match the corresponding 39-column extract in persona_records.zip.

Income and consumption labels b1–b5 are decoded with field-specific CNY interval boundaries, not a shared amount scale. Flow fields retain their annual/reference-period meaning. Debt labels none, d1–d11 use zero and right-closed balance intervals; zero and missing are distinct. Detailed education and school tiers are decoded by survey wave and education level. Occupation codes are matched to the supplied GB/T 6565 or census-2000 catalogs; unresolved fine codes are omitted from model inputs while supported coarse categories remain available. Missing values are never invented.

Targets cover household income, the available decision-maker work-income field, main income source, total and component consumption, housing tenure and debt balance. These are templates over observed harmonized categories, rather than verbatim original multiple-choice questions or simulated decisions. Income targets exclude all income inputs; consumption targets exclude all consumption inputs; housing tenure excludes tenure and housing expenditure; debt targets exclude debt amount and type. The financial asset, risk preference, financial literacy, investment, trust and insurance columns are missing in this extract and do not become synthetic labels. Brackets do not recover exact amounts or bank-default outcomes.

New household records are split 70/10/10/10 by a fixed hash. Equal source-ID codes within a survey share a split across waves as a conservative leakage guard; this does not establish verified longitudinal household identity. Raw IDs are replaced by hashes and kept out of model inputs. Existing CHFS/CGSS records and splits are retained as frozen shards. row_index is local to a shard; use pair_id for bilingual alignment. Documentation and audits are under metadata/cfps_clds/, and decoded bilingual profiles under personas/.

Observed occupation and education supervision

This release adds 255,491 aligned QA examples per language, using the existing CFPS/CLDS household records. The original source values, profile files and group/split assignments remain frozen. The added tasks are:

Target	Examples per language	Candidate design
Occupation major group	87,641	Seven fixed observed occupational categories.
Detailed educational attainment	109,524	Survey-specific separable categories; early CLDS pooled categories retain their original meaning.
Applicable college category / school affiliation	2,408	Early CLDS school affiliation is separate from later junior-college and bachelor-or-higher school categories.
Conditional detailed occupation code	55,918	A known fixed codebook prefix/leaf, with two to four predetermined catalog alternatives.

The original 242 question/task versions are retained; the addition has 969 versions across survey waves and fixed occupational candidate nodes, giving 1,211 versions in total. These versions are not 969 independent banking scenarios. Fine occupation supervision is conditional on the explicitly provided catalog prefix/leaf. The fixed candidates are generated from the supplied complete catalogs before any respondent answer is selected; no answer-dependent distractors are manufactured. This measures classification inside the known branch, rather than unconditional exact occupation prediction.

Only observed, applicable college answers become college-category targets. The 32,409 records coded zero mean below tertiary education / not applicable, and are excluded from this target; they are not an “other university” class. CFPS has no observed college-category answers. CLDS 2011/2012 asks school affiliation, while later categories depend on junior-college versus bachelor-or-higher applicability. These eligibility conditions are stated explicitly in the question. Early CLDS education codes 5 and 10 preserve vocational/technical and postgraduate pooled meanings; no missing distinctions are reconstructed.

Occupation targets remove occupational group, title, classification level and employment status from the profile. Education and school-category targets remove coarse education, detailed attainment, school category and field of study. Existing preserved coarse/fine disagreements are recorded in the source audit and cannot leak through an alias. Unknown values and unverified fine codes are not imputed. Every added language row passed label/alias checks and a lossless Parquet/native round trip. The actual exported question heads passed the deployed multilingual and English tokenizers: option spans remain distinct, all supervised option targets are trainable, and fine-code conditions and option descriptions survive the token caps.

The addition includes 178,960 train, 25,092 dev, 24,841 calibration and 26,598 test examples per language. Methods, fixed candidate groups and validation reports are in attribute-task metadata and preparation scripts. Five new stems, two conditional templates and 128 additional catalog titles were translated/reviewed by GPT-6 Luna with max reasoning effort, then applied deterministically with the existing bilingual glossary. Model training remains a separate step from this data release.

Codes, option meanings and training targets

Survey codes serve different purposes. Original variables such as a4003 and L21 identify the source questions; the model receives the full question instructions and option descriptions. Profile categories are decoded into readable values, such as education and marital status, while age and monetary amounts retain their numeric values and field units. Choice IDs such as "1" or "5" identify options within a question; custom labels such as "b" can play the same role when explicitly paired with an option description.

For example, the CHFS 2011 risk-preference item maps "5" to “Unwilling to take any risk.” An expected answer of "5" identifies that option. Laya renders each choice as ID: description, scores the candidates, and converts the observed answer ID into a target over the corresponding option slots. The returned ID can be mapped back to its description for display. The ID alone does not carry the financial meaning.

The expanded SFT and RLCD training code randomly permutes choice order and permutes the target to match. This reduces dependence on a fixed option position; it does not randomly rename option IDs. During translation and inference, keep each ID, description and observed-answer mapping consistent. Chinese and English versions preserve the same option IDs and answer targets. Missing profile values remain explicitly marked as not provided rather than introducing unexplained missing-value codes.

No dedicated label-renaming ablation has been run, so code invariance and its effect on accuracy have not been measured. A suitable follow-up check is to rename option IDs and permute their order while preserving their meanings and targets, then compare predictions after mapping back to the original options. Implementation references are included in preparation/pinned_laya_common.py (render_options), preparation/pinned_laya_train.py (target_from_expected, encode_item) and preparation/run_expanded_experiment.py (choice-order randomization).

English translation

GPT-6 Luna, max reasoning effort (gpt-6-luna, max) translated and semantically reviewed the question wording, options, profile-field names and category texts. Question-specific overrides handle ambiguous scale labels. The reviewed glossary was applied offline to every record.

Character volume is an auxiliary measure of readable text, reported in the statistics table above. It counts profile-field names and string values, instructions and option descriptions as they recur in QA records, including whitespace. JSON syntax, separately typed numeric field values, IDs, answer labels and metadata are excluded. Numbers inside text strings remain included. See metadata/text_volume_report.json.

Amounts, rates, units, time periods, negation, option IDs/order, observed answers and group assignments are preserved. Full-record validation is recorded in metadata/translation_report.json for CHFS/CGSS and metadata/cfps_clds/translation_report.json for CFPS/CLDS financial-state tasks, with the added attribute-task checks in metadata/cfps_clds_attributes/translation_report.json. Independent human bilingual review is pending.

Splits

Split	Chinese rows	English rows
train	2,925,263	2,925,263
dev	412,873	412,873
calibration	415,688	415,688
test	418,597	418,597
Total per language	4,172,421	4,172,421

Load with Hugging Face Datasets

import json
from datasets import load_dataset

english = load_dataset("ya9919/fin-laya-decision-qa", "en", split="train", streaming=True, token=False)
chinese = load_dataset("ya9919/fin-laya-decision-qa", "zh", split="test", streaming=True, token=False)
row = next(iter(english))
request = {"state": json.loads(row["state_json"]),
           "questions": json.loads(row["questions_json"])}
target = json.loads(row["expected_json"])

The examples use streaming. For full materialization, set cache_dir to a directory with sufficient free space. English is the default configuration. state_json, questions_json, expected_json and meta_json reconstruct the native decision-training record exactly. Other columns expose the question ID, prompt, option IDs/texts, observed answer, source, group and a language-independent pair_id. Use state and questions for inference; keep expected as the supervised target.

Native training files are also available under native/en/ and native/zh/ as lossless .jsonl.gz files. Each split comprises frozen CHFS/CGSS and CFPS/CLDS household-state shards, plus an added occupation/education shard. tools/materialize_native.py downloads and decompresses a selected language/split, checking the original uncompressed SHA-256. The combined training manifest is data_manifest.json, with unified question_registry_en.json and question_registry_zh.json. Component methods, translation review and export checks are included under metadata/ and preparation/.

Related models and access

ya9919/fin-laya-multilingual · ya9919/fin-laya-multilingual-mlx

Project code. This is a public research dataset. Original survey data terms remain applicable; see source terms. The dataset carries observed survey behavior, with financial and related social questions, rather than a measurement of bank defaults or credit losses.