zieglerd

RussianConcsExt

README

License: apache-2.0

Model Details

Model Description

Russian Contract Extraction LoRA is a domain-specific LoRA adapter for extracting structured information from Russian-language procurement and corporate contracts.

The adapter is fine-tuned on top of Qwen2.5-14B-Instruct and is designed to convert free-form contract text into a structured JSON representation suitable for downstream processing in enterprise document management systems, ERP platforms, procurement systems, and legal automation pipelines.

Unlike generic information extraction models, this adapter focuses on complex real-world Russian contracts, including framework agreements, contracts without fixed amounts, contracts with indirect subject descriptions, advance payment schemes, and various VAT configurations.

Developed by: Dmitry Ziegler (Lead), Denis Karpov (MLOps), Andrey Tolstov (jun. MLOps)
Model type: LoRA Adapter
Language(s): Russian
License: Apache-2.0 (inherits base model license)
Finetuned from model: Qwen/Qwen2.5-14B-Instruct

Uses

Direct Use

The adapter is intended for automated extraction of structured information from Russian contracts. The current version is optimized for contracts with a single supplier. Support for multi-supplier and multi-party agreements has not been evaluated and may require additional fine-tuning.

Input:

free-form Russian contract text;
OCR output;
document text with appendices;
specifications;
procurement documentation.

Output:

Structured JSON describing:

contract subject;
contract amount;
VAT information;
payment schedule;
advance payments;
specification items;
financial conditions.

Downstream Use

The model can be integrated into:

ERP systems;
document management systems;
procurement automation;
legal document processing;
enterprise AI assistants;
contract analytics platforms.

Out-of-Scope Use

The model is not intended to:

provide legal advice;
interpret legislation;
replace professional legal review;
validate legal compliance.

Bias, Risks and Limitations

The model is intended exclusively for information extraction.

Performance depends on:

OCR quality;
document formatting;
contract complexity;
ambiguity of legal language.

Users should validate extracted information before using it in legal or financial processes.

Training Procedure

The adapter was trained using supervised instruction tuning (SFT).

Input consists of free-form Russian procurement contract text.

Output consists of structured JSON following a fixed extraction schema.

The model was optimized to preserve factual correctness and return null whenever required information is absent instead of hallucinating unsupported values.

Training Configuration

Table with columns: Parameter, Value
Parameter	Value
Base model	Qwen/Qwen2.5-14B-Instruct
Training method	Supervised Fine-Tuning (SFT)
Adapter type	LoRA
LoRA rank	32
LoRA alpha	64
LoRA dropout	0.03
Trainable modules	`q_proj`, `k_proj`, `v_proj`, , , ,

Benchmark

The model was evaluated on an internal validation set containing Russian enterprise procurement contracts that were not used during training.

Evaluation focused on structured information extraction rather than natural language generation.

The benchmark includes contracts with:

framework agreements;
contracts without fixed total amounts;
contracts with multiple payment stages;
contracts containing specifications;
contracts with multiple VAT rates;
advance payment contracts;
contracts with complex financial conditions;
contracts containing incomplete or ambiguous information.

| Field Accuracy | 96.2% |

Table with columns: Field, Accuracy
Field	Accuracy
Contract number	99.8%
Contract date	99.3%
Supplier	98.7%
Customer	98.5%
Subject	96.1%
Contract amount	95.6%
VAT	98.9%
Payment terms	91.8%
Validity period	94.2%

Evaluation criteria

The following aspects were manually verified:

correct extraction of contract metadata;
supplier and customer identification;
subject extraction;
total contract amount extraction;
VAT identification;
payment schedule extraction;
contract validity dates;
preservation of missing values using null;
JSON schema validity;
absence of hallucinated values.

The model was optimized for high factual precision and schema consistency on long Russian legal documents (up to 32k tokens).

Evaluation

Benchmark

Russian Purchase Contracts Benchmark

Metric

Extraction Accuracy

Results

Table with columns: Model, Accuracy
Model	Accuracy
Qwen2.5-14B-Instruct	~85%

The benchmark evaluates extraction of structured information from Russian procurement contracts.

Example

Input

text
ДОГОВОР ПОСТАВКИ №458/26

ООО "Поставщик" обязуется поставить ООО "Заказчик"
общехозяйственные товары согласно спецификации.

Стоимость договора составляет
34 132 686,08 рублей,
в том числе НДС 5%.

Оплата производится следующим образом:

85% — авансовый платеж.

15% — после поставки товара.

Output

json
{
  "document_object": "Общехозяйственные товары",
  "total_amount": {
    "total": {
      "amount_value": 34132686.08,
      "vat_rate": 5,
      "include_vat": true
    }
  },
  "advance_payments": [
    {
      "percent": 85
    },
    {
      "percent": 15
    }
  ]
}

Features

The adapter supports extraction of:

contract subject;
contract amount;
VAT amount and rate;
VAT inclusion;
payment schedule;
advance payments;
post-payment terms;
specification tables;
product positions;
quantities;
unit prices;
total prices.

Special attention was given to difficult enterprise scenarios:

framework agreements;
contracts without a fixed subject;
contracts without a fixed total amount;
multiple specifications;
multiple VAT rates;
complex payment schedules.

Technical Specifications

Architecture

Base model: Qwen2.5-14B-Instruct
Adapter type: LoRA (PEFT)

Output Format

The model generates structured JSON suitable for automated processing.

Missing values are represented by null instead of inferred values, improving reliability in production environments.

Framework Versions

PEFT 0.15.1
Transformers

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

zieglerd

Model Tree

Base

Qwen/Qwen2.5-14B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Model Details

Model Description

Russian Contract Extraction LoRA is a domain-specific LoRA adapter for extracting structured information from Russian-language procurement and corporate contracts.

Developed by: Dmitry Ziegler (Lead), Denis Karpov (MLOps), Andrey Tolstov (jun. MLOps)
Model type: LoRA Adapter
Language(s): Russian
License: Apache-2.0 (inherits base model license)
Finetuned from model: Qwen/Qwen2.5-14B-Instruct

Uses

Direct Use

Input:

free-form Russian contract text;
OCR output;
document text with appendices;
specifications;
procurement documentation.

Output:

Structured JSON describing:

contract subject;
contract amount;
VAT information;
payment schedule;
advance payments;
specification items;
financial conditions.

Downstream Use

The model can be integrated into:

ERP systems;
document management systems;
procurement automation;
legal document processing;
enterprise AI assistants;
contract analytics platforms.

Out-of-Scope Use

The model is not intended to:

provide legal advice;
interpret legislation;
replace professional legal review;
validate legal compliance.

Bias, Risks and Limitations

The model is intended exclusively for information extraction.

Performance depends on:

OCR quality;
document formatting;
contract complexity;
ambiguity of legal language.

Users should validate extracted information before using it in legal or financial processes.

Training Procedure

The adapter was trained using supervised instruction tuning (SFT).

Input consists of free-form Russian procurement contract text.

Output consists of structured JSON following a fixed extraction schema.

The model was optimized to preserve factual correctness and return null whenever required information is absent instead of hallucinating unsupported values.

Training Configuration

Table with columns: Parameter, Value
Parameter	Value
Base model	Qwen/Qwen2.5-14B-Instruct
Training method	Supervised Fine-Tuning (SFT)
Adapter type	LoRA
LoRA rank	32
LoRA alpha	64
LoRA dropout	0.03
Trainable modules	`q_proj`, `k_proj`, `v_proj`, , , ,

Benchmark

The model was evaluated on an internal validation set containing Russian enterprise procurement contracts that were not used during training.

Evaluation focused on structured information extraction rather than natural language generation.

The benchmark includes contracts with:

framework agreements;
contracts without fixed total amounts;
contracts with multiple payment stages;
contracts containing specifications;
contracts with multiple VAT rates;
advance payment contracts;
contracts with complex financial conditions;
contracts containing incomplete or ambiguous information.

| Field Accuracy | 96.2% |

Table with columns: Field, Accuracy
Field	Accuracy
Contract number	99.8%
Contract date	99.3%
Supplier	98.7%
Customer	98.5%
Subject	96.1%
Contract amount	95.6%
VAT	98.9%
Payment terms	91.8%
Validity period	94.2%

Evaluation criteria

The following aspects were manually verified:

correct extraction of contract metadata;
supplier and customer identification;
subject extraction;
total contract amount extraction;
VAT identification;
payment schedule extraction;
contract validity dates;
preservation of missing values using null;
JSON schema validity;
absence of hallucinated values.

The model was optimized for high factual precision and schema consistency on long Russian legal documents (up to 32k tokens).

Evaluation

Benchmark

Russian Purchase Contracts Benchmark

Metric

Extraction Accuracy

Results

Table with columns: Model, Accuracy
Model	Accuracy
Qwen2.5-14B-Instruct	~85%

The benchmark evaluates extraction of structured information from Russian procurement contracts.

Example

Input

text
ДОГОВОР ПОСТАВКИ №458/26

ООО "Поставщик" обязуется поставить ООО "Заказчик"
общехозяйственные товары согласно спецификации.

Стоимость договора составляет
34 132 686,08 рублей,
в том числе НДС 5%.

Оплата производится следующим образом:

85% — авансовый платеж.

15% — после поставки товара.

Output

json
{
  "document_object": "Общехозяйственные товары",
  "total_amount": {
    "total": {
      "amount_value": 34132686.08,
      "vat_rate": 5,
      "include_vat": true
    }
  },
  "advance_payments": [
    {
      "percent": 85
    },
    {
      "percent": 15
    }
  ]
}

Features

The adapter supports extraction of:

contract subject;
contract amount;
VAT amount and rate;
VAT inclusion;
payment schedule;
advance payments;
post-payment terms;
specification tables;
product positions;
quantities;
unit prices;
total prices.

Special attention was given to difficult enterprise scenarios:

framework agreements;
contracts without a fixed subject;
contracts without a fixed total amount;
multiple specifications;
multiple VAT rates;
complex payment schedules.

Technical Specifications

Architecture

Base model: Qwen2.5-14B-Instruct
Adapter type: LoRA (PEFT)

Output Format

The model generates structured JSON suitable for automated processing.

Missing values are represented by null instead of inferred values, improving reliability in production environments.

Framework Versions

PEFT 0.15.1
Transformers