DarkhanOmirbay/ainn-final

stars

commits

Jupyter Notebook

primary language

May 11, 2026

updated

README

Variant 12 — Fine-tuning LLM for Business Letters (Russian)

Subject: Artificial Intelligence and Neural Networks
Group: AAI-2502M
Authors: Darhan Omirbay, Alisher Khairullin

Overview

Fine-tuned RuGPT-3 Large (760M parameters) on a synthetic corpus of 5,500 Russian business letter pairs using LoRA via transformers + peft. The model generates formal business letter replies from an incoming letter + instruction prompt.

Project Structure

variant12/
├── ainn_final_project (1).ipynb   # Full pipeline: data gen → training → eval
├── app.py                          # Streamlit demo app
├── data/
│   ├── train.json                  # 5,000 training pairs
│   ├── val.json                    # 250 validation pairs
│   └── test.json                   # 250 test pairs
├── lora_adapter/                   # Trained LoRA adapter (5k dataset)
├── lora_adapter_500/               # Experimental adapter (500 samples)
└── pyproject.toml

Dataset

5,500 synthetic Alpaca-style pairs generated with Python templates.

Split	Size
Train	5,000
Val	250
Test	250

5 letter categories (equal distribution):

запрос_информации — information requests
жалоба — complaints
коммерческое_предложение — commercial proposals
уведомление — notifications
согласование — approvals

3 response styles per category: вежливое · краткое · настойчивое

Format (Alpaca-style):

{
  "instruction": "Составьте вежливый профессиональный ответ...",
  "input": "Уважаемые коллеги,\nПрошу предоставить...",
  "output": "Уважаемый(-ая) А.В. Петрова,\n\nБлагодарим...",
  "category": "запрос_информации",
  "style": "вежливое"
}

Model & Training

Base model: ai-forever/rugpt3large_based_on_gpt2 (760M params)
Method: LoRA (FP16) via peft + transformers
Training time: 71.7 minutes (Apple MPS)

LoRA Configuration

Parameter	Value
`lora_r`	8
`lora_alpha`	16
`lora_dropout`	0.05
Target modules	`c_attn`
Trainable params	1,179,648 (0.15%) of 761M

Training Hyperparameters

Parameter	Value
Epochs	3
Learning rate	2×10⁻⁴
Batch size (effective)	16 (4 × grad_accum=4)
Max sequence length	256
LR scheduler	cosine
Precision	FP16

How to Run

1. Install dependencies

pip install transformers peft datasets accelerate torch streamlit
# or with uv:
uv sync

2. Launch the Streamlit demo

streamlit run app.py

The app loads the LoRA adapter from ./lora_adapter/ and provides an interactive UI to generate business letter replies.

3. Reproduce training / evaluation

Open ainn_final_project (1).ipynb and run all cells sequentially.

Expert Evaluation (30 test examples)

Evaluated on 30 samples from data/test.json covering all category × style combinations. Scoring is automated via a formal-marker rubric (see notebook).

Criterion	Mean	Scale
Style (деловой стиль)	4.50	1–5
Relevance (релевантность)	4.57	1–5
No hallucinations (отсутствие галлюцинаций)	4.77	1–5
Overall	4.61	1–5

Complaints (жалоба) scored slightly lower (assertive emotional tone is harder to model with a 760M parameter base).

Inference Speed

Measured with 10 greedy-decoding runs on a fixed benchmark letter.

Metric	Value
Avg latency	~5–15 sec (CPU/MPS)
Tokens/sec	measured in notebook

Run the benchmark cell in the notebook to get exact numbers for your hardware.

Libraries

References

Contributors

DarkhanOmirbay

2 commits

Similar repos

evilfreelancer/ruGPT-3.5-13B-lora

Model

wwewtech/russian-it-community-lora

Model

ai-forever/ru-gpts

Russian GPT3 models.

2,087

sberbank-ai/ru-gpts

Russian GPT3 models.

2,086

taishan1994/ChatGLM-LoRA-Tuning

使用LoRA对ChatGLM进行微调。

beyondguo/LLM-Tuning

Tuning LLMs with no tears💦; Sample Design Engineering (SDE) for more efficient downstream-tuning.

1,014

tuggspeedman-ai/SmolLM3-3B-summarize-sft-lora

Model

wwewtech/russian-it-community-corpus

Russian IT Community Conversational Corpus (2017-2026) · Zero-PII Curation Platform, Multi-turn…

Languages

Jupyter Notebook

77.9%

Python

11.3%

TeX

10.8%

DarkhanOmirbay/ainn-final

stars

commits

Jupyter Notebook

primary language

May 11, 2026

updated

README

Variant 12 — Fine-tuning LLM for Business Letters (Russian)

Subject: Artificial Intelligence and Neural Networks
Group: AAI-2502M
Authors: Darhan Omirbay, Alisher Khairullin

Overview

Project Structure

variant12/
├── ainn_final_project (1).ipynb   # Full pipeline: data gen → training → eval
├── app.py                          # Streamlit demo app
├── data/
│   ├── train.json                  # 5,000 training pairs
│   ├── val.json                    # 250 validation pairs
│   └── test.json                   # 250 test pairs
├── lora_adapter/                   # Trained LoRA adapter (5k dataset)
├── lora_adapter_500/               # Experimental adapter (500 samples)
└── pyproject.toml

Dataset

5,500 synthetic Alpaca-style pairs generated with Python templates.

Split	Size
Train	5,000
Val	250
Test	250

5 letter categories (equal distribution):

запрос_информации — information requests
жалоба — complaints
коммерческое_предложение — commercial proposals
уведомление — notifications
согласование — approvals

3 response styles per category: вежливое · краткое · настойчивое

Format (Alpaca-style):

{
  "instruction": "Составьте вежливый профессиональный ответ...",
  "input": "Уважаемые коллеги,\nПрошу предоставить...",
  "output": "Уважаемый(-ая) А.В. Петрова,\n\nБлагодарим...",
  "category": "запрос_информации",
  "style": "вежливое"
}

Model & Training

Base model: ai-forever/rugpt3large_based_on_gpt2 (760M params)
Method: LoRA (FP16) via peft + transformers
Training time: 71.7 minutes (Apple MPS)

LoRA Configuration

Parameter	Value
`lora_r`	8
`lora_alpha`	16
`lora_dropout`	0.05
Target modules	`c_attn`
Trainable params	1,179,648 (0.15%) of 761M

Training Hyperparameters

Parameter	Value
Epochs	3
Learning rate	2×10⁻⁴
Batch size (effective)	16 (4 × grad_accum=4)
Max sequence length	256
LR scheduler	cosine
Precision	FP16

How to Run

1. Install dependencies

pip install transformers peft datasets accelerate torch streamlit
# or with uv:
uv sync

2. Launch the Streamlit demo

streamlit run app.py

The app loads the LoRA adapter from ./lora_adapter/ and provides an interactive UI to generate business letter replies.

3. Reproduce training / evaluation

Open ainn_final_project (1).ipynb and run all cells sequentially.

Expert Evaluation (30 test examples)

Evaluated on 30 samples from data/test.json covering all category × style combinations. Scoring is automated via a formal-marker rubric (see notebook).

Criterion	Mean	Scale
Style (деловой стиль)	4.50	1–5
Relevance (релевантность)	4.57	1–5
No hallucinations (отсутствие галлюцинаций)	4.77	1–5
Overall	4.61	1–5

Complaints (жалоба) scored slightly lower (assertive emotional tone is harder to model with a 760M parameter base).

Inference Speed

Measured with 10 greedy-decoding runs on a fixed benchmark letter.

Metric	Value
Avg latency	~5–15 sec (CPU/MPS)
Tokens/sec	measured in notebook

Run the benchmark cell in the notebook to get exact numbers for your hardware.

Libraries

References

Similar repos

evilfreelancer/ruGPT-3.5-13B-lora

Model

wwewtech/russian-it-community-lora

Model

ai-forever/ru-gpts

Russian GPT3 models.

2,087

sberbank-ai/ru-gpts

Russian GPT3 models.

2,086

taishan1994/ChatGLM-LoRA-Tuning

使用LoRA对ChatGLM进行微调。

beyondguo/LLM-Tuning

Tuning LLMs with no tears💦; Sample Design Engineering (SDE) for more efficient downstream-tuning.

1,014

tuggspeedman-ai/SmolLM3-3B-summarize-sft-lora

Model

wwewtech/russian-it-community-corpus

Russian IT Community Conversational Corpus (2017-2026) · Zero-PII Curation Platform, Multi-turn…

Contributors

DarkhanOmirbay

2 commits

Languages

Jupyter Notebook

77.9%

Python

11.3%

TeX

10.8%