COA Mapper v1
The first open model for chart-of-accounts mapping: given a ledger account from QuickBooks Desktop or Sage 50 (name + type), retrieve the correct target account in the client's new QuickBooks Online chart of accounts.
A 33M-parameter embedding model (BAAI/bge-small-en-v1.5, fine-tuned) —
small enough to run on CPU, in a browser, or inside a migration tool.
Results — COA-Map-Bench v1
Test split of gyaanbyte/coa-mapping-synthetic:
137 mappings from 4 held-out companies (2 QuickBooks Desktop, 2 Sage 50) never
seen in training. Candidates = the distinct target accounts of the same company
(i.e. mapping into the client's new QBO COA). A top-1 hit implies the correct
QBO type and detail type.
| Method | acc@1 | acc@5 | MRR |
|---|---|---|---|
| TF-IDF (char 2–4-grams, cosine) | 0.745 | 0.949 | 0.824 |
bge-small-en-v1.5 zero-shot | 0.796 | 0.956 | 0.868 |
| qwen3:8b zero-shot (picks from the candidate list) | 0.788 | — | — |
| coa-mapper v1 (this model) | 0.883 | 0.964 | 0.924 |
Read the numbers honestly: the benchmark is synthetic, and account names carry strong lexical signal (TF-IDF already gets 74.5%). The fine-tune's gain lives in the hard 25% — abbreviations, near-duplicates to consolidate, dumping-ground splits, type/detail-type disambiguation. An adjudicated benchmark subset is planned for v1.1.
Try it
🧪 Live demo — runs in your browser via Transformers.js (int8 ONNX included in this repo under onnx/; nothing is uploaded).
Usage
from sentence_transformers import SentenceTransformer
m = SentenceTransformer("gyaanbyte/coa-mapper")
# Source accounts as "name (type) [system]"; targets as "name — type / detail type"
src = ["Accum Depr - Equipment (fixed asset) [quickbooks desktop]"]
targets = [
"Accumulated Depreciation — Fixed Assets / Accumulated Depreciation",
"Equipment — Fixed Assets / Machinery & Equipment",
"Depreciation Expense — Other Expense / Depreciation",
]
S = m.encode(src, normalize_embeddings=True)
T = m.encode(targets, normalize_embeddings=True)
print(targets[(S @ T.T).argmax()])
In production we use it as retrieval → LLM re-rank of the top-5 → confidence threshold → human review lane: the goal is hours saved per migration at $0 reconciliation variance, not unattended automation.
Training
- Base:
BAAI/bge-small-en-v1.5· loss: MultipleNegativesRankingLoss (in-batch negatives) - Data: 1,025 (source → target) pairs, train split of the dataset above
- 4 epochs, batch 32, lr 3e-5, fp16 — 12 seconds on one RTX 5060 laptop GPU
- Training/eval code:
coa_train.py(open)
Limitations
- Trained and evaluated on synthetic data; real ledgers are messier and more idiosyncratic
- Target side is QuickBooks Online only; sources are QuickBooks Desktop and Sage 50
- English, US-GAAP-oriented account structures
- Candidate pool must be supplied (the target COA); this model does not invent accounts
Who's building this
GyaanByte Labs — accounting data-migration specialists (6+ production migrations, 200,000+ records, $0 variance). The mapper is the engine behind our migration service; the reusable parts are published openly.