Upload README.md with huggingface_hub

README.md

@@ -1,35 +1,83 @@
+---
+language:
+  - ru
+  - en
+license: mit
+task_categories:
+  - text-generation
+  - conversational
+tags:
+  - code
+  - programming
+  - ai-developer
+  - fine-tuning
+  - instruction-tuning
+  - coding
+  - web-development
+  - debugging
+  - llm
+size_categories:
+  - n<1K
+---
+
 # AI Developer Dataset
 
-Dataset for fine-tuning an open-weight LLM into a universal AI developer assistant.
+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**.
 
 ## Statistics
 
 - **Total examples**: 68
 - **Format**: JSONL (conversational)
-- **Schema**: `{"messages":[{role,content}, ...], "metadata":{...}}`
+- **Schema**: `{"messages":[{"role":"system|user|assistant","content":"..."}, ...], "metadata":{"category","subcategory","complexity","source"}}`
+- **Splits**: train / validation / test
+
+## Schema
+
+Each line is a single JSON object:
+
+```json
+{
+  "messages": [
+    {"role": "system", "content": "Category: programming\nSubcategory: python_basics\n..."},
+    {"role": "user", "content": "Напиши асинхронный HTTP-клиент..."},
+    {"role": "assistant", "content": "Используем aiohttp для HTTP-запросов..."}
+  ],
+  "metadata": {
+    "category": "programming",
+    "subcategory": "python_advanced",
+    "complexity": "intermediate",
+    "source": "handcrafted"
+  }
+}
+```
+
+The system message encodes the example's category/subcategory/complexity, so the model learns to condition its style on task type.
 
 ## Categories
 
-| Category | Count |
-|----------|-------|
-| programming | 10 |
-| ui_ux_design | 10 |
-| web_development | 9 |
-| bots_automation | 8 |
-| ai_ml | 7 |
-| debugging | 6 |
-| backend_api_db | 6 |
-| linux_server_deploy | 6 |
-| web_search_docs | 6 |
+| Category | Examples | Target % |
+|----------|----------|----------|
+| programming | 10 | 25% |
+| web_development | 9 | 20% |
+| ui_ux_design | 10 | 15% |
+| debugging | 6 | 10% |
+| bots_automation | 8 | 10% |
+| backend_api_db | 6 | 5% |
+| linux_server_deploy | 6 | 5% |
+| ai_ml | 7 | 5% |
+| web_search_docs | 6 | 5% |
 
 ## Complexity
 
-| Level | Count |
-|-------|-------|
-| advanced | 20 |
-| beginner | 12 |
-| expert | 2 |
-| intermediate | 34 |
+| Level | Count | Target % |
+|-------|-------|----------|
+| beginner | 12 | 20% |
+| intermediate | 34 | 40% |
+| advanced | 20 | 30% |
+| expert | 2 | 10% |
 
 ## Subcategories
 
@@ -106,16 +154,72 @@ Every example passes:
 - `assistant` content length 200–25000 chars
 - `user` content non-empty
 - JSON-serializable (no encoding issues)
-- Category + subcategory + complexity tagged
+- Category + subcategory + complexity tagged in metadata
 - Deduplicated by user-content fingerprint (SHA-256, first 16 hex)
 
+## Content policy
+
+- All code is **real and runnable** (no pseudocode where actual code is requested)
+- No invented libraries or APIs
+- Real library versions (aiogram 3.x, FastAPI 0.110+, Next.js 14+, React 19, SQLAlchemy 2.0+, transformers 4.45+, etc.)
+- Russian explanations, English code/identifiers (consistent with target audience)
+
 ## Usage
 
 ```python
 from datasets import load_dataset
+
 ds = load_dataset('VIRUS374/ai-developer-dataset')
+print(ds)
+# DatasetDict({
+#     train: Dataset({{ features: ['messages', 'metadata'], num_rows: 57 }}),
+#     validation: Dataset({{ features: ['messages', 'metadata'], num_rows: 9 }}),
+#     test: Dataset({{ features: ['messages', 'metadata'], num_rows: 2 }})
+# }})
+
+# Iterate
+for example in ds['train']:
+    messages = example['messages']
+    # ... feed to tokenizer
 ```
 
+For fine-tuning with QLoRA:
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
+from peft import LoraConfig, get_peft_model
+from trl import SFTTrainer
+
+model = AutoModelForCausalLM.from_pretrained(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    load_in_4bit=True,
+    device_map="auto",
+)
+
+# Convert messages to chat-formatted text using tokenizer.apply_chat_template
+# Then use SFTTrainer with the formatted dataset
+```
+
+## Scaling plan
+
+- **v0.1 (current)**: 68 handcrafted high-quality examples — this release.
+- **v0.2**: 500 examples — synthetic generation with code validation pipeline.
+- **v0.3**: 2000 examples — RAG-assisted generation for current API/library versions.
+- **v1.0**: 20000 examples — final dataset for production fine-tuning.
+
+The generator pipeline is in `/scripts/` of the source repository. Each category lives in its own Python module (`cat_*.py`) and can be extended independently.
+
 ## License
 
-MIT
+MIT — use freely for fine-tuning open-weight models.
+
+## Citation
+
+```bibtex
+@dataset{{virus374_ai_developer_2024,
+  title  = {{AI Developer Dataset}},
+  author = {{VIRUS374}},
+  year   = {{2024}},
+  url    = {{https://huggingface.co/datasets/VIRUS374/ai-developer-dataset}}
+}}
+```