68standard/webdev-agentic-coder-distill-dora

WebDev Agentic Coder & Autocomplete Dataset (DoRA Ready)

Curated, filtered, and distilled dataset specifically designed for fine-tuning LLMs with DoRA (Weight-Decomposed Low-Rank Adaptation) and LoRA for:

1. Web Development: HTML5, CSS3, JavaScript (ES6+), TypeScript, React, responsive UI, DOM, and Web APIs.
2. Agentic Coding: Autonomous multi-turn problem solving, tool use (read_file, write_file, edit_file, run_command), and codebase refactoring.
3. AI Autocomplete: Fill-In-The-Middle (FIM) format (<|fim_prefix|>, <|fim_suffix|>, <|fim_middle|>) for IDE code completions.
4. Distilled Reasoning (<think> Mode): Deep chain-of-thought analysis inspired by Fable / Claude-Code traces before generating code or calling tools.
5. Bilingual (>= 40% Natural Russian): Real-world colloquial yet technical Russian paired with high-quality English.

---

Dataset Statistics

• Total Curated Examples: 3707
• Language Distribution:
  • Russian (ru): 2016 (54.4%)
  • English (en): 1691 (45.6%)
• Task Types:
  • web_reasoning_sft: 2200
  • autocomplete_fim: 1126
  • agentic_web: 381
• Reasoning Chains (<think>): 2500 examples (67.4%)
• Quality Filtering Gate: 100% passed strict heuristic filtering (>80% garbage confidence rejected).

---

Data Format

Each row adheres to a unified schema supporting both conversational training (messages) and causal completion / FIM (text):

{
  "id": "web-frontend-00042",
  "task_type": "web_reasoning_sft",
  "language": "ru",
  "framework": "javascript",
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "<think>\n...\n</think>\n..."}
  ],
  "text": "<|im_start|>system\n...<|im_end|>...",
  "meta": {
    "source": "...",
    "has_think": true,
    "has_tools": false,
    "quality_score": 0.98
  }
}

Autocomplete (FIM) Format

FIM examples use standard tokens compatible with Qwen2.5-Coder and DeepSeek-Coder:

<|fim_prefix|>...code before cursor...<|fim_suffix|>...code after cursor...<|fim_middle|>...predicted completion...

---

Fine-tuning with DoRA (Example Config)

This dataset is pre-formatted for direct ingestion by TRL / SFTTrainer, Unsloth, Axolotl, and LLaMA-Factory:

from transformers import TrainingArguments
from peft import LoraConfig
from trl import SFTTrainer
from datasets import load_dataset

dataset = load_dataset("68standard/webdev-agentic-coder-distill-dora")

peft_config = LoraConfig(
    use_dora=True,        # Enable DoRA (Weight-Decomposed LoRA)
    r=32,
    lora_alpha=64,
    lora_dropout=0.05,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    bias="none",
    task_type="CAUSAL_LM"
)

trainer = SFTTrainer(
    model="Qwen/Qwen2.5-Coder-7B-Instruct",
    train_dataset=dataset["train"],
    dataset_text_field="text",
    max_seq_length=4096,
    peft_config=peft_config,
    args=TrainingArguments(
        output_dir="./dora-web-agent-coder",
        learning_rate=2e-4,
        warmup_ratio=0.03,
        lr_scheduler_type="cosine",
        bf16=True,
        logging_steps=10,
        per_device_train_batch_size=2,
        gradient_accumulation_steps=8
    )
)
trainer.train()