surogate/ytd-dataset

YTD Dataset

Romanian fișă analitică data for training a small LM to extract Year-to-Date recipes as structured JSON.

Used to train surogate/qwen3.5-2b-libra-ytd.

Why this dataset exists

Romanian accounting software (Saga, Mentor, ContaPlus, Smartbill, custom Excel exports) emits the same trial-balance data in a dozen incompatible layouts. The YTD figure a bookkeeper actually needs is rarely a single column: it depends on the report period, on which columns the software exposes, and on which side of the ledger the account belongs to. Generic LLMs trained on English business data consistently pick the wrong column on real Romanian fișe, in particular by selecting trap columns like Sume totale (which silently include the opening balance and produce wrong YTD numbers).

We built this dataset to teach a small model the rule, not the surface form. Each training sample renders the same underlying scenario through one of 12 surface formats (PDF copy, Excel paste, OCR-mangled, headerless, fixed-width, markdown, etc.), with 50% probability of having the helper section markers stripped so the model must infer period and account root from raw text. The output schema is a recipe (formula + column references), not a number, so a deterministic post-processor can apply the recipe and the model never has to do arithmetic.

How it teaches generalization

signal in data	what the model learns
12 surface formats per pattern	column header text is not load-bearing; structural position is
50% marker dropout	period and cont rădăcină inferred from raw header text, not the helper line
Pattern F renders both F4 and F5 GTs	subtraction and composition are interchangeable when both column families exist
Trap-column samples (Sume totale, Rulaj+SI) with explicit avoidance	model picks Rulaje perioada even when a more prominent trap column is present
14 real production fișe replicated ×30 as anchors	shapes the model toward real-world spacing and tokenization patterns
not_applicable samples for 404/461/5121/4426 roots	refusal behavior is preserved instead of fabricating a 411/401 answer

Splits

split	records	source
train	3620	3200 synth across 6 layouts + 420 real-fișă anchors (14 docs ×30)
validation	320	independent generation, seed 99999, 0 verbatim overlap with train

Schema

field	type	content
instruction	str	fixed Romanian system prompt with the decision rules
input	str	raw extracted ledger text (PDF copy, Excel paste, OCR, CSV, markdown, etc.)
response	str	JSON string with ytd_debit, ytd_credit, optional tip_column, or { "not_applicable": true }

Response JSON shape

field	values
ytd_debit.formula	ytd_source | ytd_source - sold_initial | addend1 + addend2
ytd_debit.ytd_source	{ header, index }
ytd_debit.sold_initial	only for subtractive
ytd_debit.addend1, addend2	only for composed
ytd_credit	same shape as ytd_debit
tip_column	{ header, index } when partner-type column exists
not_applicable	true for non-411/401 account roots

Patterns (train split)

pattern	layout	count	example program
A	Layout 1 single-period "Rulaje perioada"	600	Uzina Brașov
B	Layout 3' "Sume precedente" trap + "Rulaje perioada"	600	Andisol
C	Layout 3'' headerless variant	400	Moise
D	Layout 2 "Total rulaje" present	600	Bughero
E	Layout 5 with "Tip" partner-type column	400	Skyline
F	Layout 4/4' single-month requires subtraction or composition	600	(Caz 1 specialty)

Pattern F splits 60% F4 (subtractive) / 40% F5 (composed). F4 and F5 are mathematically equivalent when both column sets exist.

Formula mix

formula	train	%	val	%
ytd_source (direct)	2960	81.8	260	81.3
addend1 + addend2 (composed)	386	10.7	41	12.8
ytd_source - sold_initial (subtractive)	274	7.6	19	5.9

Input formats

Each sample is rendered through one of 12 surface formats (weighted-random pick per sample):

format	description
canonical	clean structured
excel_paste	TSV-like
pdf_copy	column-misaligned
csv	comma-separated
markdown_table	pipe-separated with header rule
fixed_width	space-padded columns
multi_line_header	header wraps to 2 rows
cont_prefix	account code prefixed to row
ocr_mangled	character substitutions, joined cells
minimal	whitespace-only separators
verbose	extra meta columns
mixed_language	Romanian + English headers

Marker dropout

split	section markers present
train	1798 / 3620 (50.3%)
validation	129 / 320 (40.3%)

Section markers are PERIOADĂ DETECTATĂ:, CONT RĂDĂCINĂ DETECTAT:, COLOANE (N total):, FIȘĂ ANALITICĂ. When stripped, the model must infer them from raw text.

Decision rules (encoded in instruction)

input signal	output
period starts 01.01	Caz 2, direct ytd_source from YTD column
single month + "Total rulaje" column	Caz 1, direct ytd_source
single month + "Perioada precedentă" + "Perioada"	Caz 1, addend1 + addend2
single month + "Sume totale" / "Rulaj+SI"	Caz 1, ytd_source - sold_initial
cont rădăcină starts 411	side = Debit
cont rădăcină starts 401	side = Credit
cont rădăcină 404 / 461 / 5121 / 4426 / etc.	not_applicable

Trap columns to never pick directly as ytd_source: Sume totale, Total sume, Rulaj+SI, Sume precedente (all include opening balance).

Trained model results

eval	score
Real production fișe all-axes	14/14 strict
Column indices on indexed fișe	12/12
Held-out synthetic all-axes	88%

Load

from datasets import load_dataset
ds = load_dataset("surogate/ytd-dataset")
# DatasetDict({ train: 3620, validation: 320 })
sample = ds["train"][0]

Generation

step	source
pattern scenarios	datagen/gen_v16_scenarios.py
12 surface renderers	datagen/v16_render.py
marker dropout (50%)	datagen/gen_v23a_marker_dropout.py
assembly	datagen/gen_v23b_realworld.py
train seed	23123
validation seed	99999

License

Apache 2.0. Synthetic data plus 14 anonymized real-fișă layout anchors.