Zenith-7b-V1 / data /preprocessing.py
Zandy-Wandy's picture
Upload Zenith-7B model
8d18b7c verified
Raw
History Blame Contribute Delete
10.6 kB
"""Advanced Preprocessing for OpenThoughts and Custom Datasets"""
import json
import logging
import re
from typing import Any, Dict, List, Optional, Tuple
logger = logging.getLogger(__name__)
# Special token markers
THOUGHT_START = "<think>"
THOUGHT_END = "</think>"
USER_START = "<user>"
USER_END = "</user>"
ASSISTANT_START = "<assistant>"
ASSISTANT_END = "</assistant>"
SYSTEM_START = "<system>"
SYSTEM_END = "</system>"
def preprocess_conversation(
conversations: Any,
include_thoughts: bool = True,
include_reasoning: bool = True,
) -> Dict[str, Any]:
"""Preprocess conversation data into training format."""
if isinstance(conversations, str):
try:
conversations = json.loads(conversations)
except json.JSONDecodeError:
return {"text": conversations, "conversations": []}
if not isinstance(conversations, list):
return {"text": str(conversations), "conversations": []}
processed_messages = []
thoughts = []
reasoning = ""
for msg in conversations:
if not isinstance(msg, dict):
continue
role = msg.get("role", "").lower()
content = msg.get("content", "")
if not content:
continue
# Extract thoughts if present
if include_thoughts and THOUGHT_START in content:
thought_parts = re.findall(r'<think>(.*?)</think>', content, re.DOTALL)
thoughts.extend([t.strip() for t in thought_parts if t.strip()])
# Remove thought tags from content
content = re.sub(r'<think>.*?</think>', '', content, flags=re.DOTALL).strip()
# Format with special tokens
if role == "user":
formatted = f"{USER_START} {content} {USER_END}"
elif role == "assistant":
formatted = f"{ASSISTANT_START} {content} {ASSISTANT_END}"
elif role == "system":
formatted = f"{SYSTEM_START} {content} {SYSTEM_END}"
else:
formatted = content
processed_messages.append({
"role": role,
"content": content,
"formatted": formatted,
})
# Combine into single text
text = "\n".join(msg["formatted"] for msg in processed_messages)
result = {
"text": text,
"conversations": processed_messages,
}
if include_thoughts and thoughts:
result["thoughts"] = " ".join(thoughts)
if include_reasoning and reasoning:
result["reasoning"] = reasoning
return result
def extract_thoughts(text: str) -> str:
"""Extract chain-of-thought from text."""
pattern = re.compile(r'<think>(.*?)</think>', re.DOTALL)
thoughts = pattern.findall(text)
return " ".join(t.strip() for t in thoughts if t.strip())
def format_for_training(
sample: Dict[str, Any],
include_thoughts: bool = True,
include_reasoning: bool = True,
) -> str:
"""Format sample for model training."""
if "text" in sample:
text = sample["text"]
elif "conversations" in sample:
text = preprocess_conversation(sample["conversations"], include_thoughts, include_reasoning)["text"]
elif "content" in sample:
text = sample["content"]
else:
text = ""
# Add thoughts if available and requested
if include_thoughts and "thoughts" in sample and sample["thoughts"]:
text += f"\n{THOUGHT_START} {sample['thoughts']} {THOUGHT_END}"
return text
def detect_domain(conversations: Any) -> str:
"""Detect domain of conversation based on content."""
if isinstance(conversations, str):
try:
conversations = json.loads(conversations)
except:
conversations = []
text = ""
for msg in conversations:
if isinstance(msg, dict):
text += msg.get("content", "") + " "
text = text.lower()
# Domain keywords
domain_keywords = {
"code": ["def ", "class ", "import ", "function", "return", "if __name__", "```python", "```java", "```cpp"],
"mathematics": ["equation", "theorem", "proof", "calculate", "solve", "integral", "derivative", "matrix", "vector"],
"science": ["experiment", "hypothesis", "theory", "data", "analysis", "chemical", "physical", "biological"],
"reasoning": ["because", "therefore", "thus", "hence", "since", "logic", "deduce", "infer"],
"dialogue": ["how are you", "what do you think", "please help", "thank you", "could you"],
}
scores = {}
for domain, keywords in domain_keywords.items():
score = sum(1 for kw in keywords if kw in text)
scores[domain] = score
if not scores:
return "unknown"
return max(scores, key=scores.get)
def estimate_difficulty(conversations: Any, thoughts: str = "") -> float:
"""Estimate difficulty on scale 0-1."""
if isinstance(conversations, str):
try:
conversations = json.loads(conversations)
except:
conversations = []
text = ""
for msg in conversations:
if isinstance(msg, dict):
text += msg.get("content", "") + " "
text += thoughts
# Features for difficulty
features = {
"length": len(text.split()),
"technical_terms": len(re.findall(r'\b[A-Z][a-z]+(?:[A-Z][a-z]+)+\b', text)), # CamelCase
"code_blocks": len(re.findall(r'```[\s\S]*?```', text)),
"math_symbols": len(re.findall[r'[=+\-*/<>≤≥≠∈∉⊂⊆∪∩]', text]),
"reasoning_markers": len(re.findall(r'\b(because|therefore|thus|hence|since)\b', text, re.IGNORECASE)),
}
# Normalize and combine
difficulty = (
min(features["length"] / 500, 1.0) * 0.3 +
min(features["technical_terms"] / 20, 1.0) * 0.25 +
min(features["code_blocks"] / 3, 1.0) * 0.25 +
min(features["math_symbols"] / 10, 1.0) * 0.1 +
min(features["reasoning_markers"] / 5, 1.0) * 0.1
)
return min(difficulty, 1.0)
def clean_text(text: str) -> str:
"""Clean and normalize text."""
# Remove excessive whitespace
text = re.sub(r'\s+', ' ', text)
# Remove control characters
text = re.sub(r'[\x00-\x1F\x7F]', '', text)
# Normalize quotes
text = text.replace('"', '"').replace('"', '"')
text = text.replace(''', "'").replace(''', "'")
# Strip
text = text.strip()
return text
def truncate_with_overlap(
text: str,
max_length: int,
stride: int,
tokenizer: Any,
) -> List[Dict[str, Any]]:
"""Truncate long text with overlapping windows."""
tokens = tokenizer.encode(text, add_special_tokens=False)
if len(tokens) <= max_length:
return [{"input_ids": tokens, "attention_mask": [1] * len(tokens)}]
chunks = []
start = 0
while start < len(tokens):
end = min(start + max_length, len(tokens))
chunk_tokens = tokens[start:end]
chunks.append({
"input_ids": chunk_tokens,
"attention_mask": [1] * len(chunk_tokens),
})
if end >= len(tokens):
break
start += stride
return chunks
def compute_length_statistics(lengths: List[int]) -> Dict[str, float]:
"""Compute statistics for length distribution."""
import numpy as np
if not lengths:
return {}
arr = np.array(lengths)
return {
"mean": float(np.mean(arr)),
"std": float(np.std(arr)),
"min": float(np.min(arr)),
"max": float(np.max(arr)),
"p50": float(np.percentile(arr, 50)),
"p90": float(np.percentile(arr, 90)),
"p95": float(np.percentile(arr, 95)),
"p99": float(np.percentile(arr, 99)),
}
def analyze_dataset_quality(dataset: Any, sample_size: int = 1000) -> Dict[str, Any]:
"""Analyze dataset quality metrics."""
logger.info("Analyzing dataset quality...")
# Sample dataset
if hasattr(dataset, "__len__"):
sample_size = min(sample_size, len(dataset))
indices = list(range(sample_size))
else:
# Streaming dataset
samples = []
for i, sample in enumerate(dataset):
if i >= sample_size:
break
samples.append(sample)
dataset = samples
sample_size = len(samples)
analysis = {
"total_samples": sample_size,
"domains": {},
"difficulty_distribution": {},
"length_stats": {},
"thoughts_coverage": 0.0,
"conversation_turns": [],
}
domains = []
difficulties = []
lengths = []
thoughts_counts = []
turns = []
for sample in dataset:
# Domain
domain = sample.get("domain", detect_domain(sample.get("conversations", [])))
domains.append(domain)
# Difficulty
difficulty = sample.get("difficulty", estimate_difficulty(sample.get("conversations", []), sample.get("thoughts", "")))
difficulties.append(difficulty)
# Length
text = sample.get("text", "")
if not text and "conversations" in sample:
text = preprocess_conversation(sample["conversations"])["text"]
lengths.append(len(text.split()))
# Thoughts
if "thoughts" in sample and sample["thoughts"]:
thoughts_counts.append(1)
else:
thoughts_counts.append(0)
# Turns
if "conversations" in sample and isinstance(sample["conversations"], list):
turns.append(len(sample["conversations"]))
# Compute statistics
from collections import Counter
analysis["domains"] = dict(Counter(domains))
analysis["difficulty_distribution"] = {
"mean": float(np.mean(difficulties)) if difficulties else 0.0,
"std": float(np.std(difficulties)) if difficulties else 0.0,
"histogram": np.histogram(difficulties, bins=10, range=(0, 1))[0].tolist(),
}
analysis["length_stats"] = compute_length_statistics(lengths)
analysis["thoughts_coverage"] = sum(thoughts_counts) / len(thoughts_counts) if thoughts_counts else 0.0
analysis["conversation_turns"] = {
"mean": float(np.mean(turns)) if turns else 0.0,
"max": int(max(turns)) if turns else 0,
}
logger.info(f"Dataset analysis complete: {analysis}")
return analysis