surogate/mf-dataset

MF Dataset (registre de mijloace fixe)

Romanian registru de mijloace fixe (fixed-asset register) data for training a small LM to emit a column-mapping recipe as structured JSON. Used to train surogate/Qwen3-1.7B-Libra-MF.

Why this dataset exists

Romanian accounting software (Soft1, Saga, Mentor, SmartBill, custom Excel exports) emits the same fixed-asset register in a dozen incompatible layouts. For each 3-digit asset category an accountant needs six totals: entry value, improvements, inventory value, accumulated depreciation, monthly depreciation, net book value. But which column holds each varies by software, headers differ, and the register comes in two structural shapes (section-grouped vs. per-row Cont column). Generic LLMs pick the wrong column (monthly vs. cumulative depreciation, a bare Valoare trap, an integer decoy).

We built this dataset to teach the model structural position, not header surface form. Each sample renders one underlying register scenario through one of 12 surface formats, and the model's target is a recipe ({role: {header, index}}), not the numbers; a deterministic post-processor applies the recipe, so the model never does arithmetic.

How it teaches generalization

signal in data	what the model learns
12 surface formats per scenario (column order preserved)	header text is not load-bearing; structural position is
recipe header = the header as rendered (English / OCR-mangled / wrapped)	copy the surface header, so the post-processor can anchor it
headerless format up-weighted (~21 % of mix)	identify value columns by their numeric pattern / position, emitting header: ""
optional value columns dropped (modernizări/amortizare lunară/rămasă)	emit null for an absent role instead of hallucinating one
integer decoy columns (Durata funct., Luni rămase) + bare Valoare	avoid trap columns; never map a duration as money
both grouped (cont = null) and column (cont present) shapes	infer the register shape and where the category comes from
7 real client registers replicated ×50 as anchors	real-world spacing / tokenization patterns

Splits

split	records	source
train	5,750	5,400 synthetic + 350 real-register anchors (7 docs × 50)
validation	600	held-out split of the same generation

Every record is token-budgeted to ≤ 2048 tokens (no truncation): the input is windowed to the header + as many sample rows as fit, since column mapping needs only a header and a few rows.

Schema

field	type	content
instruction	str	fixed Romanian system prompt with the role definitions and rules (~1,340 chars)
input	str	raw register text in one of 12 surface formats
response	str	JSON recipe: { "coloane": { <role>: {header, index} | null } }

Response JSON shape

{ "coloane": {
  "denumire":            {"header": "...", "index": 1},
  "valoare_intrare":     {"header": "...", "index": 2},
  "valoare_modernizari": {"header": "...", "index": 4},
  "valoare_inventar":    {"header": "...", "index": 5},
  "valoare_amortizata":  {"header": "...", "index": 6},
  "amortizare_luna":     {"header": "...", "index": 7},
  "valoare_ramasa":      {"header": "...", "index": 8},
  "cont":                {"header": "...", "index": 10}
} }

Any role's value may be null if that column is absent. cont = null marks a grouped register (category comes from section headers); cont present marks a column register.

Asset categories

Standard Romanian 3-digit fixed-asset accounts, balanced across the set:

code	category
205	Concesiuni, brevete, licențe, mărci
208	Alte imobilizări necorporale
211	Terenuri (no depreciation)
212	Construcții
213	Instalații tehnice și mijloace de transport
214	Mobilier, aparatură birotică
215	Investiții imobiliare

Archetypes in train: grouped 3,297 / column 2,703.

Surface formats

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

format	description	count
headerless	data rows only, no header row (position-only)	1,261
canonical_markers	structured COLOANE [i] / RÂNDURI [i] markers	630
cont_prefix	account code bulleted per row	448
fixed_width	space-padded columns	435
pdf_copy	single-space-collapsed (pdfplumber extract_text)	426
multi_line_header	header wrapped across two rows	410
markdown	pipe table with header rule	409
mixed_language	Romanian + English headers	407
ocr_mangled	letter substitutions (digits intact), pipe-separated	400
csv	comma-separated (quote-aware)	394
tsv	tab-separated	390
header_only	header row + data, no markers	390

Number styles vary per sample: ro (2.219,31), plain (1805972.00), space (1 805 972.00).

Trained model results

eval	score
7 real client registers (end-to-end 6-field totals)	7 / 7
360 held-out synthetic, all formats	95.3 %
600 validation (recipe agreement)	92.6 %

Load

from datasets import load_dataset
ds = load_dataset("surogate/mf-dataset")   # DatasetDict({ train: 5750, validation: 600 })
sample = ds["train"][0]

Generation

step	source
scenario builders (grouped / column)	generate_mf.py (build_grouped, build_column)
12 surface renderers	mf_surface.py (render)
recipe target (surface headers)	generate_mf.py (build_label) + mf_surface.surface_headers
token-budget windowing (≤2048)	generate_mf.py (render_input)
real-register anchors (×50)	anchors.json
deterministic post-processor (recipe → totals)	mf_apply.py + mf_executor.py
train seed	42

License

Apache 2.0. Synthetic data plus 7 anonymized real-register layout anchors.