distill::synthetic
More...
Classes
Attributes
Detailed Description
Synthetic data generation using multi-backend cascade-capable teacher models.```
## Attributes Documentation
### variable QUALITY_MIN_CHARS
```python
int QUALITY_MIN_CHARS = 200;
variable REFUSAL_PATTERNS
list REFUSAL_PATTERNS = [
r"\bI cannot\b",
r"\bI['\u2019]m unable\b",
r"\bas an AI\b",
r"\bI don['\u2019]t have\b",
r"\bI do not have\b",
r"\bI am not able\b",
r"\bI['\u2019]m not able\b",
r"\bsorry.*cannot\b",
r"\bcan['\u2019]t (?:help|assist|do that|generate|create|provide)\b",
];
variable parser
parser = argparse.ArgumentParser(description="Generate synthetic data for a specialist niche");
variable required
variable True
variable help
variable args
args = parser.parse_args();
variable project_root
project_root = Path(__file__).resolve().parent.parent;
variable loader
loader = ConfigLoader(project_root);
variable cfg
cfg = loader.get_effective_config(args.niche);
variable system_prompt
system_prompt = cfg.get("system_prompt", f"You are a {args.niche} specialist.");
variable user_prompts
user_prompts = cfg.get("synthetic_prompts", [f"Explain {args.niche} concepts in detail."]);
variable client
client = TeacherClient(project_root);
variable generator
generator = SyntheticDataGenerator(client, project_root, use_cascade=True);
variable samples
samples = generator.generate_for_niche(args.niche, system_prompt, user_prompts);
Updated on 2026-07-25 at 22:56:57 +0000