patent-wireframes / scripts /eval /condition_c_text_only.py
midah's picture
Reorganize: scripts/eval/condition_c_text_only.py
520af59 verified
"""Condition C: Text-only viewpoint identification — no image.
The four conditions being tested across the benchmark:
A — Vision only: VLM sees image, no text
B — Drawing program: LLM sees SVG path tokens, no image (future)
C — Text description: LLM sees drawing description + numeric features, no image
D — Vision + text: VLM sees image AND drawing description (future)
This script runs Condition C.
TASK 1 — Viewpoint Identification:
Given a representation of a single patent figure, name its viewpoint type
(e.g. "front elevational view", "perspective view", "top plan view").
Ground truth: viewpoint label parsed from drawing_description text.
Scoring: loose keyword overlap (same as Vision condition A).
Current Vision (A) result: 23.3% overall, 5.3% on front elevational.
TASK 1 Condition C setup:
Input: patent title + full drawing_description with the TARGET FIGURE's viewpoint
word MASKED (replaced with ___). The model sees what all OTHER figures show, but
must predict what the target figure's viewpoint is.
This is FIM (Fill-in-the-Middle) on the drawing description text — exactly
analogous to code FIM where prefix/suffix constrain the middle.
Numeric features (ink_frac, edge_transitions, img_aspect) are also provided
as a supplementary signal when available.
TASK 2 — Cross-view Correspondence (text version):
Given drawing descriptions of N-1 views of a patent, plus a candidate set of
view descriptions (with viewpoint labels masked), identify which candidate is
the front elevational view.
Currently deferred — Task 2 is broken with the thinking model in the vision
condition; running text version here would not be a fair comparison.
Usage:
python scripts/eval/condition_c_text_only.py \
--enriched data/enriched/enriched_2022.parquet \
--sample data/sample/eval_sample.parquet \
--n 18 \
--out results/condition_c_results.json
"""
import argparse
import json
import re
import time
from collections import defaultdict
from pathlib import Path
import pandas as pd
import sys
sys.path.insert(0, str(Path(__file__).parent))
from provider import chat, get_client
# ── viewpoint parsing (shared with track_a) ──────────────────────────────────
def parse_viewpoint(desc: str, fig_num: int) -> str:
pat = re.compile(
rf"FIG\.\s*{fig_num + 1}\s+is\s+(?:a\s+|an\s+)?(.{{5,80}}?)\s*(?:view|thereof|;|\n|$)",
re.IGNORECASE,
)
m = pat.search(desc or "")
return m.group(1).strip().lower() if m else ""
def mask_viewpoint(desc: str, fig_num: int) -> tuple[str, str]:
"""Replace the target figure's viewpoint with ___ and return (masked_desc, original_vp)."""
pat = re.compile(
rf"(FIG\.\s*{fig_num + 1}\s+is\s+(?:a\s+|an\s+)?)(.{{5,80}}?)(\s*(?:view|thereof|;|\n|$))",
re.IGNORECASE,
)
vp = parse_viewpoint(desc, fig_num)
masked = pat.sub(r"\1___\3", desc or "", count=1)
return masked, vp
def viewpoint_match(predicted: str, ground_truth: str) -> bool:
DIRECTIONS = {"front","rear","back","left","right","top","bottom","side",
"perspective","plan","elevation","elevational","isometric",
"oblique","reference","detail","enlarged","cross","section"}
pred_w = set(re.findall(r"\w+", predicted.lower())) & DIRECTIONS
gt_w = set(re.findall(r"\w+", ground_truth.lower())) & DIRECTIONS
if not gt_w:
return False
return len(pred_w & gt_w) / len(gt_w) >= 0.5
VP_BUCKETS = [
("perspective", lambda v: "perspective" in v),
("front_elev", lambda v: "front elev" in v),
("rear_elev", lambda v: "rear elev" in v or "rear elevation" in v),
("side_elev", lambda v: "side elev" in v or "side elevation" in v),
("top_plan", lambda v: "top plan" in v or v == "top view"),
("bottom_plan", lambda v: "bottom plan" in v or "bottom" in v),
("other", lambda v: True),
]
def bucket(vp: str) -> str:
for name, fn in VP_BUCKETS:
if fn(vp):
return name
return "other"
# ── prompt construction ───────────────────────────────────────────────────────
TASK1_C_PROMPT = """\
You are an engineer reading a US design patent.
Patent title: {title}
The patent has {n_figs} figures. Here is the drawing description with one figure's viewpoint type replaced by ___:
{masked_desc}
{numeric_context}
What type of view is FIG. {fig_num}? Give the viewpoint label only (2-5 words).
Examples: "front elevational view", "perspective view", "top plan view", "rear elevational view"
"""
def make_prompt(row: pd.Series) -> str | None:
"""Build the text-only Task 1 prompt for one figure."""
desc = str(row.get("drawing_description") or "")
fig_num = int(row.get("figure_number", 0))
title = str(row.get("patent_title") or "")
n_figs = int(row.get("n_figures_in_patent", 0))
masked_desc, vp = mask_viewpoint(desc, fig_num)
if not vp:
return None, None # can't evaluate without ground truth
# Numeric context (if available)
numeric_parts = []
for col, label in [("ink_frac","ink density"), ("edge_transitions","edge transitions"),
("img_aspect","aspect ratio")]:
val = row.get(col)
if val is not None and not pd.isna(val):
if col == "ink_frac":
numeric_parts.append(f" {label}: {float(val):.2%}")
elif col == "edge_transitions":
numeric_parts.append(f" {label}: {int(val):,}")
else:
numeric_parts.append(f" {label}: {float(val):.2f}")
numeric_context = ("Image statistics for FIG. {}:\n{}".format(fig_num + 1, "\n".join(numeric_parts))
if numeric_parts else "")
prompt = TASK1_C_PROMPT.format(
title=title,
n_figs=n_figs,
masked_desc=masked_desc.strip(),
numeric_context=numeric_context,
fig_num=fig_num + 1,
)
return prompt, vp
# ── main ──────────────────────────────────────────────────────────────────────
def run(enriched_path: str, sample_path: str | None, n: int, out_path: str, seed: int = 42):
client = get_client()
df = pd.read_parquet(enriched_path)
# If sample parquet provided, restrict to those patent_ids
if sample_path and Path(sample_path).exists():
sample = pd.read_parquet(sample_path)
df = df[df["patent_id"].isin(sample["patent_id"].unique())]
print(f"Restricted to {df['patent_id'].nunique()} sample patents")
# Pick same eligible patents as track_a (perspective + front + ≥3 figs)
df["vp"] = df.apply(lambda r: parse_viewpoint(r.get("drawing_description",""), r["figure_number"]), axis=1)
has_persp = df.groupby("patent_id")["vp"].apply(lambda x: x.str.contains("perspective").any())
has_front = df.groupby("patent_id")["vp"].apply(lambda x: (x.str.contains("front elev") | x.isin(["front elevation","front plan"])).any())
eligible = has_persp.index[has_persp & has_front].tolist()
import random
rng = random.Random(seed)
rng.shuffle(eligible)
eval_pids = eligible[:n]
print(f"Eligible patents: {len(eligible)}, evaluating: {len(eval_pids)}")
results = []
by_vp = defaultdict(lambda: [0, 0]) # [correct, total]
total_correct = total_n = 0
for pid in eval_pids:
group = df[df["patent_id"] == pid].sort_values("figure_number")
patent_results = []
for _, row in group.iterrows():
prompt, ground_truth = make_prompt(row)
if prompt is None:
continue
# ── TASK 1 (Condition C): text-only viewpoint identification ──────
# Input: masked drawing_description (FIM on text) + numeric features
# Output: predicted viewpoint type
# No image is shown — this is a pure language model inference task
msgs = [{"role": "user", "content": prompt}]
predicted = chat(client, msgs, max_tokens=30).lower().strip()
correct = viewpoint_match(predicted, ground_truth)
vp_label = bucket(ground_truth)
by_vp[vp_label][1] += 1
if correct:
by_vp[vp_label][0] += 1
total_correct += int(correct)
total_n += 1
patent_results.append({
"fig": int(row["figure_number"]),
"ground_truth": ground_truth,
"predicted": predicted,
"correct": correct,
"vp_bucket": vp_label,
})
time.sleep(0.3)
if patent_results:
n_correct = sum(r["correct"] for r in patent_results)
results.append({
"patent_id": pid,
"patent_title": str(group["patent_title"].iloc[0]),
"figures": patent_results,
})
print(f"[{pid}] {str(group['patent_title'].iloc[0])[:45]} — "
f"{n_correct}/{len(patent_results)} correct")
# ── Summary ───────────────────────────────────────────────────────────────
print()
print("=" * 60)
print("CONDITION C — TEXT-ONLY TASK 1 RESULTS")
print("Input: drawing description (FIM, viewpoint masked) + numeric features")
print("No image shown")
print("=" * 60)
print(f"Overall: {total_correct}/{total_n} = {total_correct/max(total_n,1):.1%}")
print()
print(f"{'Viewpoint':<16} {'Correct':>8} {'Total':>6} {'Acc':>6}")
for label, (c, t) in sorted(by_vp.items(), key=lambda x: -x[1][1]):
if t == 0: continue
print(f" {label:<14} {c:>8} {t:>6} {c/t:>5.1%}")
print()
print("COMPARISON TABLE")
print(f" Condition A (vision only): 23.3% overall | 5.3% front elev")
print(f" Condition C (text only): {total_correct/max(total_n,1):>5.1%} overall |"
f" {by_vp['front_elev'][0]/max(by_vp['front_elev'][1],1):>4.1%} front elev")
print(f" Human baseline (est): ~95%")
output = {
"condition": "C_text_only",
"task": "Task 1 — Viewpoint Identification",
"input": "drawing_description (FIM: target viewpoint masked) + numeric features",
"no_image": True,
"summary": {
"overall_acc": total_correct / max(total_n, 1),
"n": total_n,
"by_viewpoint": {k: {"correct": v[0], "total": v[1], "acc": v[0]/v[1] if v[1] else 0}
for k, v in by_vp.items()},
},
"results": results,
"comparison": {
"condition_A_vision_only": {"overall": 0.233, "front_elev": 0.053},
"condition_C_text_only": {"overall": total_correct/max(total_n,1),
"front_elev": by_vp["front_elev"][0]/max(by_vp["front_elev"][1],1)},
}
}
Path(out_path).parent.mkdir(parents=True, exist_ok=True)
with open(out_path, "w") as f:
json.dump(output, f, indent=2)
print(f"\nResults → {out_path}")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--enriched", default="data/enriched/enriched_2022.parquet")
parser.add_argument("--sample", default="data/sample/eval_sample.parquet")
parser.add_argument("--n", type=int, default=18)
parser.add_argument("--out", default="results/condition_c_results.json")
args = parser.parse_args()
run(args.enriched, args.sample, args.n, args.out)
if __name__ == "__main__":
main()