---
language:
  - en
  - ru
license:
  - mit
  - apache-2.0
  - cc-by-4.0
  - cc-by-sa-4.0
  - odc-by
task_categories:
  - text-generation
tags:
  - code
  - programming
  - ai-developer
  - fine-tuning
  - instruction-tuning
  - coding
  - web-development
  - debugging
  - llm
  - security
  - sql
  - python
  - javascript
  - typescript
  - react
size_categories:
  - 1M<n<10M
---

# AI Developer Dataset

A large-scale instruction-tuning dataset for fine-tuning an open-weight LLM into a **universal AI developer assistant**.

The model trained on this dataset should be especially good at:
**PROGRAMMING + WEB DEVELOPMENT + UI/UX + ANIMATIONS + BOTS + SCRIPTS + AUTOMATION + BACKEND + API + DATABASES + DEBUGGING + LINUX + DEPLOYMENT + AI DEVELOPMENT + SECURITY**.

## 📊 Statistics

- **Total examples**: 1,124,699
- **File size**: 2.13 GB
- **Format**: JSONL (conversational)
- **Average assistant length**: ~1,222 characters
- **Schema**: `{"messages":[{"role":"system|user|assistant","content":"..."}, ...], "metadata":{"category","subcategory","complexity","source"}}`

## 📚 Sources

This dataset is a curated merge of high-quality open-source code instruction datasets from HuggingFace, plus handcrafted examples in Russian covering UI/UX, security, and other specialized topics.

| Source | Examples |
|--------|----------|
| `hf:glaiveai/glaive-code-assistant-v2` | 213,670 |
| `hf:glaiveai/glaive-code-assistant` | 135,195 |
| `hf:mwitiderrick/glaive-code-assistant` | 104,336 |
| `hf:b-mc2/sql-create-context` | 78,573 |
| `hf:TokenBender/code_instructions_122k_alpaca_style` | 78,433 |
| `hf:iamtarun/code_instructions_120k_alpaca` | 78,433 |
| `hf:ajibawa-2023/Code-74k-ShareGPT` | 68,790 |
| `hf:Locutusque/code-feedback-sharegpt` | 62,866 |
| `hf:m-a-p/Code-Feedback` | 62,866 |
| `hf:HuggingFaceH4/Code-Feedback` | 61,923 |
| `hf:WizardLMTeam/WizardLM_evol_instruct_70k` | 60,411 |
| `hf:nickrosh/Evol-Instruct-Code-80k-v1` | 58,265 |
| `hf:Data-Agora/magpie_code_gpt4o_mini_50000` | 36,818 |
| `hf:ajibawa-2023/Python-Code-23k-ShareGPT` | 17,737 |
| `hf:HuggingFaceH4/CodeAlpaca_20K` | 6,307 |
| `handcrafted` | 76 |


## 📁 Categories

| Category | Examples | % |
|----------|----------|---|
| programming | 1,046,060 | 93.0% |
| backend_api_db | 78,579 | 7.0% |
| ui_ux_design | 10 | 0.0% |
| web_development | 9 | 0.0% |
| bots_automation | 8 | 0.0% |
| security | 8 | 0.0% |
| ai_ml | 7 | 0.0% |
| debugging | 6 | 0.0% |
| linux_server_deploy | 6 | 0.0% |
| web_search_docs | 6 | 0.0% |


## 🎯 Complexity Distribution

| Level | Examples | % |
|-------|----------|---|
| intermediate | 686,519 | 61.0% |
| advanced | 438,166 | 39.0% |
| beginner | 12 | 0.0% |
| expert | 2 | 0.0% |


## 📋 Schema

Each line is a single JSON object:

```json
{
  "messages": [
    {"role": "system", "content": "Category: programming\nSubcategory: code_alpaca\n..."},
    {"role": "user", "content": "Write a Python function to..."},
    {"role": "assistant", "content": "Here is the implementation..."}
  ],
  "metadata": {
    "category": "programming",
    "subcategory": "code_alpaca",
    "complexity": "intermediate",
    "source": "hf:HuggingFaceH4/CodeAlpaca_20K"
  }
}
```

## 🛠️ Quality Control

Every example passes:
- JSON-serializable (validated on 100% of examples)
- Non-empty user and assistant content
- Assistant length between 200 and 25,000 characters
- Per-source deduplication by user-content fingerprint (SHA-256, first 16 hex chars)
- Schema validation (messages + metadata required)

## 🚀 Usage

```python
from datasets import load_dataset

# Load the full dataset
ds = load_dataset("VIRUS374/ai-developer-dataset", data_files="dataset.jsonl", split="train")
print(ds)
# Dataset({
#     features: ["messages", "metadata"],
#     num_rows: 1,124,699
# })

# Or use streaming for memory-efficient iteration
ds = load_dataset("VIRUS374/ai-developer-dataset", data_files="dataset.jsonl", split="train", streaming=True)
for example in ds:
    messages = example["messages"]
    # ... feed to tokenizer with chat template
```

## 🏋️ Fine-tuning Recipe (QLoRA)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from trl import SFTTrainer
from datasets import load_dataset
import torch

MODEL = "meta-llama/Llama-3.1-8B-Instruct"

tokenizer = AutoTokenizer.from_pretrained(MODEL)
tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    MODEL,
    torch_dtype=torch.bfloat16,
    load_in_4bit=True,
    device_map="auto",
)
model = prepare_model_for_kbit_training(model)

lora_config = LoraConfig(
    r=16, lora_alpha=32,
    target_modules=["q_proj", "v_proj", "k_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    lora_dropout=0.05, bias="none", task_type="CAUSAL_LM",
)
model = get_peft_model(model, lora_config)

ds = load_dataset("VIRUS374/ai-developer-dataset", data_files="dataset.jsonl", split="train")

def format_example(example):
    text = tokenizer.apply_chat_template(
        example["messages"], tokenize=False, add_generation_prompt=False,
    )
    return {"text": text}

ds = ds.map(format_example, remove_columns=ds.column_names)

trainer = SFTTrainer(
    model=model,
    train_dataset=ds,
    args=TrainingArguments(
        output_dir="./output",
        num_train_epochs=3,
        per_device_train_batch_size=2,
        gradient_accumulation_steps=8,
        warmup_steps=50,
        learning_rate=2e-4,
        bf16=True,
        logging_steps=10,
        save_steps=500,
    ),
    max_seq_length=4096,
)
trainer.train()
```

## 📜 License & Attribution

This dataset aggregates multiple open-source datasets. Each retains its original license:

- **HuggingFaceH4/CodeAlpaca_20K** — Apache 2.0
- **HuggingFaceH4/Code-Feedback** — Apache 2.0
- **nickrosh/Evol-Instruct-Code-80k-v1** — Apache 2.0
- **WizardLMTeam/WizardLM_evol_instruct_70k** — Apache 2.0 (Non-commercial)
- **TokenBender/code_instructions_122k_alpaca_style** — Apache 2.0
- **iamtarun/code_instructions_120k_alpaca** — Apache 2.0
- **m-a-p/Code-Feedback** — CC BY-NC 4.0
- **Locutusque/code-feedback-sharegpt** — Apache 2.0
- **ajibawa-2023/Code-74k-ShareGPT** — Apache 2.0
- **ajibawa-2023/Python-Code-23k-ShareGPT** — Apache 2.0
- **b-mc2/sql-create-context** — CC BY 4.0
- **glaiveai/glaive-code-assistant** — Apache 2.0
- **glaiveai/glaive-code-assistant-v2** — Apache 2.0
- **mwitiderrick/glaive-code-assistant** — Apache 2.0
- **Data-Agora/magpie_code_gpt4o_mini_50000** — Apache 2.0
- **Handcrafted examples** (VIRUS374) — MIT

**Important:** If you use this dataset for commercial purposes, please review the WizardLM and m-a-p licenses — they may have non-commercial restrictions.

## 🙏 Acknowledgements

Thanks to all the original dataset creators:
- HuggingFace H4 team
- WizardLM team
- TokenBender
- iamtarun
- m-a-p
- Locutusque
- ajibawa-2023
- b-mc2
- glaiveai
- mwitiderrick
- Data-Agora

## 📝 Citation

```bibtex
@dataset{virus374_ai_developer_2024,
  title  = {AI Developer Dataset},
  author = {VIRUS374},
  year   = {2024},
  url    = {https://huggingface.co/datasets/VIRUS374/ai-developer-dataset},
  note   = {Aggregated from open-source code instruction datasets + handcrafted examples}
}
```