68standard commited on
Commit
2da25d7
·
verified ·
1 Parent(s): 706b454

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +125 -0
README.md ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ task_categories:
4
+ - text-generation
5
+ language:
6
+ - ru
7
+ - en
8
+ tags:
9
+ - web-development
10
+ - html
11
+ - css
12
+ - javascript
13
+ - typescript
14
+ - react
15
+ - agentic-coding
16
+ - autocomplete
17
+ - fim
18
+ - dora
19
+ - reasoning
20
+ - fable-distillation
21
+ pretty_name: "WebDev Agentic Coder & Autocomplete with Distilled Reasoning (DoRA Ready)"
22
+ size_categories:
23
+ - 1K<n<10K
24
+ ---
25
+
26
+ # WebDev Agentic Coder & Autocomplete Dataset (DoRA Ready)
27
+
28
+ **Curated, filtered, and distilled dataset** specifically designed for fine-tuning LLMs with **DoRA** (Weight-Decomposed Low-Rank Adaptation) and LoRA for:
29
+ 1. **Web Development**: HTML5, CSS3, JavaScript (ES6+), TypeScript, React, responsive UI, DOM, and Web APIs.
30
+ 2. **Agentic Coding**: Autonomous multi-turn problem solving, tool use (`read_file`, `write_file`, `edit_file`, `run_command`), and codebase refactoring.
31
+ 3. **AI Autocomplete**: Fill-In-The-Middle (FIM) format (`<|fim_prefix|>`, `<|fim_suffix|>`, `<|fim_middle|>`) for IDE code completions.
32
+ 4. **Distilled Reasoning (`<think>` Mode)**: Deep chain-of-thought analysis inspired by Fable / Claude-Code traces before generating code or calling tools.
33
+ 5. **Bilingual (>= 40% Natural Russian)**: Real-world colloquial yet technical Russian paired with high-quality English.
34
+
35
+ ---
36
+
37
+ ## Dataset Statistics
38
+
39
+ - **Total Curated Examples**: 2583
40
+ - **Language Distribution**:
41
+ - **Russian (ru)**: 1459 (56.5%)
42
+ - **English (en)**: 1124 (43.5%)
43
+ - **Task Types**:
44
+ - `web_reasoning_sft`: 2200
45
+ - `autocomplete_fim`: 0
46
+ - `agentic_web`: 383
47
+ - **Reasoning Chains (`<think>`)**: 2500 examples (96.8%)
48
+ - **Quality Filtering Gate**: 100% passed strict heuristic filtering (>80% garbage confidence rejected).
49
+
50
+ ---
51
+
52
+ ## Data Format
53
+
54
+ Each row adheres to a unified schema supporting both conversational training (`messages`) and causal completion / FIM (`text`):
55
+
56
+ ```json
57
+ {
58
+ "id": "web-frontend-00042",
59
+ "task_type": "web_reasoning_sft",
60
+ "language": "ru",
61
+ "framework": "javascript",
62
+ "messages": [
63
+ {"role": "system", "content": "..."},
64
+ {"role": "user", "content": "..."},
65
+ {"role": "assistant", "content": "<think>\n...\n</think>\n..."}
66
+ ],
67
+ "text": "<|im_start|>system\n...<|im_end|>...",
68
+ "meta": {
69
+ "source": "...",
70
+ "has_think": true,
71
+ "has_tools": false,
72
+ "quality_score": 0.98
73
+ }
74
+ }
75
+ ```
76
+
77
+ ### Autocomplete (FIM) Format
78
+ FIM examples use standard tokens compatible with Qwen2.5-Coder and DeepSeek-Coder:
79
+ ```
80
+ <|fim_prefix|>...code before cursor...<|fim_suffix|>...code after cursor...<|fim_middle|>...predicted completion...
81
+ ```
82
+
83
+ ---
84
+
85
+ ## Fine-tuning with DoRA (Example Config)
86
+
87
+ This dataset is pre-formatted for direct ingestion by **TRL / SFTTrainer**, **Unsloth**, **Axolotl**, and **LLaMA-Factory**:
88
+
89
+ ```python
90
+ from transformers import TrainingArguments
91
+ from peft import LoraConfig
92
+ from trl import SFTTrainer
93
+ from datasets import load_dataset
94
+
95
+ dataset = load_dataset("68standard/webdev-agentic-coder-distill-dora")
96
+
97
+ peft_config = LoraConfig(
98
+ use_dora=True, # Enable DoRA (Weight-Decomposed LoRA)
99
+ r=32,
100
+ lora_alpha=64,
101
+ lora_dropout=0.05,
102
+ target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
103
+ bias="none",
104
+ task_type="CAUSAL_LM"
105
+ )
106
+
107
+ trainer = SFTTrainer(
108
+ model="Qwen/Qwen2.5-Coder-7B-Instruct",
109
+ train_dataset=dataset["train"],
110
+ dataset_text_field="text",
111
+ max_seq_length=4096,
112
+ peft_config=peft_config,
113
+ args=TrainingArguments(
114
+ output_dir="./dora-web-agent-coder",
115
+ learning_rate=2e-4,
116
+ warmup_ratio=0.03,
117
+ lr_scheduler_type="cosine",
118
+ bf16=True,
119
+ logging_steps=10,
120
+ per_device_train_batch_size=2,
121
+ gradient_accumulation_steps=8
122
+ )
123
+ )
124
+ trainer.train()
125
+ ```