surogate/Qwen3.5-2B-Libra-YTD

Qwen3.5-2B-Libra-YTD

Qwen3.5-2B fine-tuned to read Romanian fișe analitice (analytical accounting ledgers) and emit Year-to-Date extraction recipes as structured JSON. LoRA SFT, then merged back into a single 2B checkpoint for drop-in inference.

Trained on `surogate/ytd-dataset`.

Business use case

Romanian SMBs report VAT and trial balances monthly. Accounting software (Saga, Mentor, Smartbill, ContaPlus, custom Excel exports) emits the same trial-balance data in a dozen incompatible layouts. A bookkeeper opening any of them needs the YTD figure per partner: how much was invoiced or paid cumulative since January.

That figure is rarely a single column. Depending on the report:

• —It is a direct read from Rulaje perioada, Rulaj curent, or Total rulaje, but only when the period covers the full year.
• —It is a subtraction Sume totale - Sume precedente when the only column shown bundles in the opening balance.
• —It is a sum Perioada precedentă + Perioada when the report breaks the year into chunks.
• —The side depends on the account root: 411x goes on Debit, 401x goes on Credit, anything else returns not_applicable.
• —A small list of trap columns (Sume totale, Total sume, Rulaj+SI, Sume precedente) looks correct but includes the opening balance and silently produces wrong numbers.

This model reads the raw extracted text (PDF copy, Excel paste, OCR, CSV, headerless, fixed-width, markdown table, multi-line header, mixed-language) and emits a JSON recipe naming exactly which column to read and which formula to apply. A deterministic post-processor consumes the recipe and produces the actual YTD figure. The model handles layout variance, the post-processor handles arithmetic.

Why we built a dedicated SML for this

Production users running GPT-4-class general models on Romanian fișe were hitting the same failure modes over and over:

Bigger general-purpose models did not solve this because the problem is not a knowledge problem. The information to disambiguate is sitting in the input every time. The problem is that those models were never trained to attend to it. A 2B model that has seen 3620 rule-teaching examples in 12 surface formats outperforms general 70B+ models on this task at a fraction of the inference cost.

How the training data makes it generalize

The same scenario is rendered through 12 surface formats per sample. 50% of training inputs have helper section markers (PERIOADĂ DETECTATĂ:, CONT RĂDĂCINĂ DETECTAT:, COLOANE) stripped so the model has to infer period and account root from raw text. Pattern F samples render the same underlying YTD value through both the F4 (subtractive) and F5 (composed) recipes, teaching the model that the two are interchangeable when both column families exist. Trap-column samples explicitly teach avoidance: even when Sume totale is the most visible column, the model picks the right one. 14 real production fișe are replicated ×30 as anchors against tokenization drift.

Eval results

Per-pattern accuracy on the 320 validation set:

Worked examples (real production fișe)

Pattern A: Uzina Brașov, full year, paired suffix headers

Input (extracted from PDF):
  PERIOADĂ DETECTATĂ: 01.01.2025 -- 31.10.2025
  CONT RĂDĂCINĂ DETECTAT: 4111
  COLOANE (10 total, 4 numerice):
    [4] Rulaje perioada Dt
    [5] Rulaje perioada Cr
    [6] Sume totale Dt          <- trap
    [7] Sume totale Cr          <- trap

{
  "ytd_debit":  { "formula": "ytd_source", "ytd_source": { "header": "Rulaje perioada Dt", "index": 4 } },
  "ytd_credit": { "formula": "ytd_source", "ytd_source": { "header": "Rulaje perioada Cr", "index": 5 } }
}

Period starts 01.01, so this is Caz 2: direct read. Model picks Rulaje perioada (4/5) and avoids the visually-similar Sume totale trap (6/7).

Pattern B: Andisol, full year, named-and-empty header pair

Input:
  PERIOADĂ DETECTATĂ: 01.01.2025 -- 31.05.2025
  CONT RĂDĂCINĂ DETECTAT: 4111
  COLOANE:
    [2] Sume precedente        <- trap (prior period total)
    [4] Rulaje perioada
    [5] (empty)
    [6] Sume totale            <- trap

{
  "ytd_debit":  { "formula": "ytd_source", "ytd_source": { "header": "",  "index": 4 } },
  "ytd_credit": { "formula": "ytd_source", "ytd_source": { "header": "",  "index": 5 } }
}

The Debit column has a header Rulaje perioada but the Credit column it spans has an empty header (the software collapses paired columns). The model correctly emits "header": "" for the empty cell rather than fabricating a name. It also skips Sume precedente (the prior-period trap) and Sume totale.

Pattern C: Moise, headerless Caz 2

Input:
  PERIOADĂ DETECTATĂ: 01.01.2024 -- 31.12.2024
  CONT RĂDĂCINĂ DETECTAT: 4111
  COLOANE (12 total, 0 numerice):
    [0] Cont
    [1] Denumirea contului
    [2..11] (all empty)
  RÂNDURI: 4111 CLIENTI 0.00 0.00 0.00 0.00 5 850 344.52 5 650 175.52 ...

{
  "ytd_debit":  { "formula": "ytd_source", "ytd_source": { "header": "", "index": 6 } },
  "ytd_credit": { "formula": "ytd_source", "ytd_source": { "header": "", "index": 7 } }
}

All headers empty; only the row data reveals column positions. The model identifies columns 6 and 7 by counting numeric cells in the data rows, not by header text. This is the layout where header-matching general models tend to fabricate plausible-looking but invented column names.

Pattern D: Bughero, layout with Total rulaje instead of Rulaje perioada

Input:
  PERIOADĂ DETECTATĂ: 01.01.2025 -- 31.12.2025
  CONT RĂDĂCINĂ DETECTAT: 4111
  COLOANE:
    [4] Rulaje perioada Debitoare       <- single period, NOT YTD here
    [5] Rulaje perioada Creditoare
    [6] Total rulaje an Debitoare       <- this is YTD
    [7] Total rulaje an Creditoare
    [8] Sume totale Debit               <- trap

{
  "ytd_debit":  { "formula": "ytd_source", "ytd_source": { "header": "Total rulaje an Debitoare", "index": 6 } },
  "ytd_credit": { "formula": "ytd_source", "ytd_source": { "header": "Total rulaje an Creditoare", "index": 7 } }
}

When Total rulaje is present in the same fișă as Rulaje perioada, Total rulaje is the YTD column. The model picks 6/7 over the more visually familiar 4/5.

Pattern E: Skyline, partner-type column present

Input:
  PERIOADĂ DETECTATĂ: Ianuarie 2025 - August 2025
  CONT RĂDĂCINĂ DETECTAT: 411
  COLOANE (7 total):
    [2] Tip                            <- partner type (Persoană fizică / juridică)
    [4] Rulaj curent | debit
    [5] Rulaj curent | credit

{
  "ytd_debit":  { "formula": "ytd_source", "ytd_source": { "header": "Rulaj curent debit",  "index": 4 } },
  "ytd_credit": { "formula": "ytd_source", "ytd_source": { "header": "Rulaj curent credit", "index": 5 } },
  "tip_column": { "header": "Tip", "index": 2 }
}

When a partner-type column exists, the model emits tip_column so the post-processor can join on partner type for VAT-split reporting.

Pattern F: Algorithm, single month (Caz 1)

Input:
  PERIOADĂ DETECTATĂ: Perioada 9         <- single month, not 01.01-...
  CONT RĂDĂCINĂ DETECTAT: 4111
  COLOANE (15 total):
    [4]  Sold initial | Debit
    [5]  Sold initial | Credit
    [6]  Perioada precedente | Debit     <- F5 addend1
    [7]  Perioada precedente | Credit
    [8]  Perioada | Debit                <- F5 addend2
    [9]  Perioada | Credit
    [10] Total sume | Debit              <- F4 source
    [11] Total sume | Credit

{
  "ytd_debit": {
    "formula": "ytd_source - sold_initial",
    "ytd_source":   { "header": "Total sume Debit",   "index": 10 },
    "sold_initial": { "header": "Sold initial Debit", "index": 4 }
  },
  "ytd_credit": {
    "formula": "ytd_source - sold_initial",
    "ytd_source":   { "header": "Total sume Credit",   "index": 11 },
    "sold_initial": { "header": "Sold initial Credit", "index": 5 }
  }
}

Perioada 9 means "month 9", not "from Jan to month 9". This is Caz 1: the YTD cannot be read directly. The model emits the subtractive recipe (F4): Total sume - Sold initial. The composed recipe (F5) would yield the same number: Perioada precedente + Perioada.

Not applicable: non-receivable/payable account

Input:
  PERIOADĂ DETECTATĂ: 01.01.2025 -- 31.12.2025
  CONT RĂDĂCINĂ DETECTAT: 5121          <- bank account, not 411/401

{ "not_applicable": true, "reason": "cont rădăcină 5121 nu este 411x sau 401x" }

When the account root is not a receivable (411x) or payable (401x), the model refuses with a structured reason rather than fabricating a YTD recipe.

Decision rules (what the model learned)

Trap columns that the model never picks as ytd_source directly: Sume totale, Total sume, Rulaj+SI, Sume precedente (all include opening balance).

Output schema

Quick start

transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("surogate/Qwen3.5-2B-Libra-YTD", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "surogate/Qwen3.5-2B-Libra-YTD",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# System prompt comes from the dataset's `instruction` field; load any sample to use it.
from datasets import load_dataset
SYSTEM = load_dataset("surogate/ytd-dataset", split="train[:1]")[0]["instruction"]

user_text = open("my_fisa.txt").read()
messages = [{"role": "user", "content": f"{SYSTEM}\n{user_text}"}]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = tok(prompt, return_tensors="pt").to(model.device)
with torch.inference_mode():
    out = model.generate(**inputs, max_new_tokens=1024, do_sample=False)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

vLLM

vllm serve surogate/Qwen3.5-2B-Libra-YTD \
    --max-model-len 4096 \
    --gpu-memory-utilization 0.85 \
    --language-model-only

curl -s http://localhost:8000/v1/chat/completions \
    -H 'Content-Type: application/json' \
    -d '{
      "model": "surogate/Qwen3.5-2B-Libra-YTD",
      "messages": [{"role": "user", "content": "<system prompt>\n<fișă text>"}],
      "temperature": 0,
      "max_tokens": 1024
    }'

Training details

The Qwen3.5-2B base uses a hybrid Gated DeltaNet + Gated Attention architecture, which has 12 distinct LoRA-targetable projection modules (not the usual 7). Training without the extra in_proj_* and out_proj targets leaves most of the model's adaptive capacity untouched.

Limitations

• —Romanian only. The mixed_language training format introduces some English headers but the model is not robust to fully English fișe.
• —Receivables (411x) and payables (401x) are the main focus. Other account roots (404, 461, 5121, 4426, ...) return not_applicable rather than a YTD recipe.
• —Inputs longer than 2048 tokens were truncated during training. Very long fișe should be passed first-N-rows-truncated at inference, or with max_model_len 4096 if the fișă fits.
• —The 14 real production anchors are replicated rather than augmented. Generalization to truly novel layouts relies on the synthetic 12-format mix; layouts unlike any of A through F are not guaranteed.
• —The post-processor that consumes the recipe is not part of this checkpoint. Without it, the model output is a recipe, not a number.

License

Apache 2.0. Inherits from Qwen/Qwen3.5-2B-Base. Synthetic training data plus 14 anonymized real-fișă layout anchors.

Citation

@misc{qwen35-2b-libra-ytd,
  title  = {Qwen3.5-2B-Libra-YTD: Romanian fișă analitică YTD recipe extractor},
  author = {Surogate},
  year   = {2026},
  url    = {https://huggingface.co/surogate/Qwen3.5-2B-Libra-YTD}
}