StepProbe / stepprobe /align.py
Akiyue's picture
Add files using upload-large-folder tool
1e59964 verified
Raw
History Blame Contribute Delete
8.05 kB
"""
StepProbe: Step Alignment
Aligns the step sequences from a full-precision model and a quantized model
for the same problem, enabling step-by-step comparison even when the two
traces have different numbers of steps or different boundaries.
Methods:
1. DTW (Dynamic Time Warping) on step embeddings
2. Longest Common Subsequence on step text
3. Index-based (simple 1:1 matching by position)
"""
import difflib
from dataclasses import dataclass
from typing import List, Tuple, Optional
import numpy as np
@dataclass
class StepAlignment:
"""A single aligned pair of steps."""
ref_index: Optional[int] # index in reference (FP16), None if insertion
hyp_index: Optional[int] # index in hypothesis (quantized), None if deletion
ref_text: str
hyp_text: str
similarity: float # 0.0 - 1.0
alignment_type: str # "match" | "substitution" | "insertion" | "deletion"
def text_similarity(a: str, b: str) -> float:
"""Compute text similarity between two step strings using SequenceMatcher."""
if not a and not b:
return 1.0
if not a or not b:
return 0.0
return difflib.SequenceMatcher(None, a.lower(), b.lower()).ratio()
def align_by_index(ref_steps: List[dict], hyp_steps: List[dict]) -> List[StepAlignment]:
"""
Simple index-based alignment: pair step i with step i.
Good when both traces have similar structure.
"""
alignments = []
max_len = max(len(ref_steps), len(hyp_steps))
for i in range(max_len):
ref = ref_steps[i] if i < len(ref_steps) else None
hyp = hyp_steps[i] if i < len(hyp_steps) else None
ref_text = ref["text"] if ref else ""
hyp_text = hyp["text"] if hyp else ""
if ref and hyp:
sim = text_similarity(ref_text, hyp_text)
atype = "match" if sim > 0.8 else "substitution"
elif ref and not hyp:
sim = 0.0
atype = "deletion"
else:
sim = 0.0
atype = "insertion"
alignments.append(StepAlignment(
ref_index=ref["index"] if ref else None,
hyp_index=hyp["index"] if hyp else None,
ref_text=ref_text,
hyp_text=hyp_text,
similarity=sim,
alignment_type=atype,
))
return alignments
def align_by_dtw(ref_steps: List[dict], hyp_steps: List[dict]) -> List[StepAlignment]:
"""
Dynamic Time Warping alignment using text similarity as cost.
Better for traces with different numbers of steps.
"""
n = len(ref_steps)
m = len(hyp_steps)
if n == 0 and m == 0:
return []
if n == 0:
return [StepAlignment(None, h["index"], "", h["text"], 0.0, "insertion") for h in hyp_steps]
if m == 0:
return [StepAlignment(r["index"], None, r["text"], "", 0.0, "deletion") for r in ref_steps]
# Compute cost matrix (1 - similarity)
cost = np.zeros((n + 1, m + 1))
cost[0, :] = np.arange(m + 1)
cost[:, 0] = np.arange(n + 1)
sim_matrix = np.zeros((n, m))
for i in range(n):
for j in range(m):
sim_matrix[i, j] = text_similarity(ref_steps[i]["text"], hyp_steps[j]["text"])
for i in range(1, n + 1):
for j in range(1, m + 1):
sub_cost = cost[i - 1, j - 1] + (1.0 - sim_matrix[i - 1, j - 1])
del_cost = cost[i - 1, j] + 1.0
ins_cost = cost[i, j - 1] + 1.0
cost[i, j] = min(sub_cost, del_cost, ins_cost)
# Traceback
alignments = []
i, j = n, m
while i > 0 or j > 0:
if i > 0 and j > 0:
sub_cost = cost[i - 1, j - 1] + (1.0 - sim_matrix[i - 1, j - 1])
del_cost = cost[i - 1, j] + 1.0
ins_cost = cost[i, j - 1] + 1.0
min_cost = min(sub_cost, del_cost, ins_cost)
if min_cost == sub_cost:
sim = sim_matrix[i - 1, j - 1]
atype = "match" if sim > 0.8 else "substitution"
alignments.append(StepAlignment(
ref_index=ref_steps[i - 1]["index"],
hyp_index=hyp_steps[j - 1]["index"],
ref_text=ref_steps[i - 1]["text"],
hyp_text=hyp_steps[j - 1]["text"],
similarity=sim,
alignment_type=atype,
))
i -= 1
j -= 1
elif min_cost == del_cost:
alignments.append(StepAlignment(
ref_index=ref_steps[i - 1]["index"],
hyp_index=None,
ref_text=ref_steps[i - 1]["text"],
hyp_text="",
similarity=0.0,
alignment_type="deletion",
))
i -= 1
else:
alignments.append(StepAlignment(
ref_index=None,
hyp_index=hyp_steps[j - 1]["index"],
ref_text="",
hyp_text=hyp_steps[j - 1]["text"],
similarity=0.0,
alignment_type="insertion",
))
j -= 1
elif i > 0:
alignments.append(StepAlignment(
ref_index=ref_steps[i - 1]["index"],
hyp_index=None,
ref_text=ref_steps[i - 1]["text"],
hyp_text="",
similarity=0.0,
alignment_type="deletion",
))
i -= 1
else:
alignments.append(StepAlignment(
ref_index=None,
hyp_index=hyp_steps[j - 1]["index"],
ref_text="",
hyp_text=hyp_steps[j - 1]["text"],
similarity=0.0,
alignment_type="insertion",
))
j -= 1
alignments.reverse()
return alignments
def align_steps(
ref_steps: List[dict],
hyp_steps: List[dict],
method: str = "dtw",
) -> List[StepAlignment]:
"""
Align reference (FP16) and hypothesis (quantized) step sequences.
Args:
ref_steps: List of step dicts from the full-precision model
hyp_steps: List of step dicts from the quantized model
method: "dtw" or "index"
Returns:
List of StepAlignment objects
"""
if method == "index":
return align_by_index(ref_steps, hyp_steps)
elif method == "dtw":
return align_by_dtw(ref_steps, hyp_steps)
else:
raise ValueError(f"Unknown alignment method: {method}. Use 'dtw' or 'index'.")
def alignment_summary(alignments: List[StepAlignment]) -> dict:
"""Compute summary statistics for an alignment."""
n = len(alignments)
if n == 0:
return {"n_aligned": 0}
types = [a.alignment_type for a in alignments]
sims = [a.similarity for a in alignments if a.alignment_type in ("match", "substitution")]
return {
"n_aligned": n,
"n_match": types.count("match"),
"n_substitution": types.count("substitution"),
"n_insertion": types.count("insertion"),
"n_deletion": types.count("deletion"),
"avg_similarity": float(np.mean(sims)) if sims else 0.0,
"min_similarity": float(np.min(sims)) if sims else 0.0,
}
def format_alignment(alignments: List[StepAlignment], max_text_len: int = 60) -> str:
"""Format an alignment for human-readable display."""
lines = []
lines.append(f"{'Type':<14} {'Ref#':<6} {'Hyp#':<6} {'Sim':<6} {'Ref Text':<{max_text_len}} {'Hyp Text'}")
lines.append("-" * (14 + 6 + 6 + 6 + max_text_len * 2 + 5))
for a in alignments:
ref_idx = str(a.ref_index) if a.ref_index is not None else "-"
hyp_idx = str(a.hyp_index) if a.hyp_index is not None else "-"
ref_t = a.ref_text[:max_text_len].replace("\n", " ")
hyp_t = a.hyp_text[:max_text_len].replace("\n", " ")
lines.append(f"{a.alignment_type:<14} {ref_idx:<6} {hyp_idx:<6} {a.similarity:<6.2f} {ref_t:<{max_text_len}} {hyp_t}")
return "\n".join(lines)