File size: 8,552 Bytes
991f3b9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 | """CP-10: Slice evaluation — analyze per-form, per-modifier, per-borderline,
per-specificity, per-industry, and length-bucket performance.
Reads predictions from work/models/preds_*.csv (HF-LLM outputs) and the trained-encoder
predictions from work/models/<model>/test_predictions_seed*.csv.
Outputs work/results/slices.json, plus heatmaps in work/figures/.
"""
from __future__ import annotations
import json, ast, glob
from pathlib import Path
import numpy as np
import pandas as pd
from sklearn.metrics import f1_score, accuracy_score
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
ROOT = Path("/home/aniket/praxis-benchmark")
WORK = ROOT / "work"
DATA = WORK / "data"
RES = WORK / "results"
FIG = WORK / "figures"
MODELS = WORK / "models"
def parse_list(s):
if isinstance(s, list): return s
if pd.isna(s): return []
try:
x = ast.literal_eval(str(s))
if isinstance(x, list): return x
except Exception: pass
return []
def load_test():
g = pd.read_csv(DATA / "gold_test.csv", low_memory=False)
g['is_suggestion'] = g['is_suggestion'].astype(str).str.lower().isin(['true','1'])
g['borderline'] = g['borderline'].astype(str).str.lower().isin(['true','1'])
g['true'] = g['is_suggestion'].astype(int)
g['tier2_list'] = g['tier2_modifiers'].apply(parse_list)
g['text'] = g['text'].fillna('').astype(str)
g['n_words'] = g['text'].str.split().apply(len)
# Use gold_id as the praxis_id (predictions are saved with this column renamed)
if 'gold_id' in g.columns and 'praxis_id' not in g.columns:
g['praxis_id'] = g['gold_id']
return g
def find_preds():
out = {}
# Encoder predictions
for f in glob.glob(str(MODELS / "*/test_predictions_seed*.csv")):
p = Path(f)
model = p.parent.name
seed = p.stem.split('seed')[-1]
try:
df = pd.read_csv(p)
out.setdefault(f"enc_{model}", []).append((seed, df))
except Exception:
pass
# HF LLM predictions
for f in glob.glob(str(MODELS / "preds_hf_*.csv")):
p = Path(f)
try:
df = pd.read_csv(p)
mname = p.stem.replace('preds_hf_', '')
out[f"llm_{mname}"] = [('zs', df)]
except Exception:
pass
return out
def macro_f1(true, pred):
return float(f1_score(true, pred, average='macro', zero_division=0))
def per_form(g, preds):
out = {}
for form in ['Direct imperative', 'Modal/deontic', 'Conditional', 'Optative',
'Interrogative', 'Comparative']:
sub = g[g['tier1_form'] == form]
if len(sub) < 5: continue
m = sub['praxis_id'].isin(preds.index) if 'praxis_id' in sub.columns else None
# join on praxis_id
merged = sub.merge(preds.reset_index(), on='praxis_id', how='inner')
if len(merged) == 0: continue
# Recall on positives of this form
recall = float(((merged['pred'] == 1) & (merged['true'] == 1)).sum() /
max((merged['true'] == 1).sum(), 1))
out[form] = {'n': int(len(merged)), 'recall': recall}
return out
def per_modifier(g, preds):
out = {}
for mod in ['Frustrated', 'Hedged', 'Backhanded', 'Sarcastic']:
mask = g['tier2_list'].apply(lambda l: mod in l)
sub = g[mask & (g['true'] == 1)]
merged = sub.merge(preds.reset_index(), on='praxis_id', how='inner')
if len(merged) < 5: continue
out[mod] = {'n': int(len(merged)),
'recall': float((merged['pred'] == 1).mean())}
return out
def per_specificity(g, preds):
out = {}
for spec in ['Vague', 'Specific', 'Very specific']:
sub = g[(g['tier3_specificity'] == spec) & (g['true'] == 1)]
merged = sub.merge(preds.reset_index(), on='praxis_id', how='inner')
if len(merged) < 5: continue
out[spec] = {'n': int(len(merged)),
'recall': float((merged['pred'] == 1).mean())}
return out
def per_borderline(g, preds):
out = {}
for label, val in [('borderline', True), ('clear', False)]:
sub = g[(g['borderline'].astype(bool) == val) & (g['true'] == 1)]
merged = sub.merge(preds.reset_index(), on='praxis_id', how='inner')
if len(merged) < 5: continue
out[label] = {'n': int(len(merged)),
'recall': float((merged['pred'] == 1).mean())}
return out
def per_macro(g, preds):
out = {}
for macro in g['macro_domain'].dropna().unique():
sub = g[g['macro_domain'] == macro]
merged = sub.merge(preds.reset_index(), on='praxis_id', how='inner')
if len(merged) < 20: continue
out[macro] = {
'n': int(len(merged)),
'macro_f1': macro_f1(merged['true'], merged['pred']),
'pos_f1': float(f1_score(merged['true'], merged['pred'], pos_label=1, zero_division=0)),
}
return out
def per_length(g, preds, n_buckets=4):
out = {}
qs = g['n_words'].quantile(np.linspace(0, 1, n_buckets + 1)).values
for i in range(n_buckets):
lo, hi = qs[i], qs[i + 1]
sub = g[(g['n_words'] >= lo) & (g['n_words'] <= hi if i == n_buckets - 1 else g['n_words'] < hi)]
merged = sub.merge(preds.reset_index(), on='praxis_id', how='inner')
if len(merged) < 20: continue
out[f'L{i+1}_[{int(lo)}-{int(hi)}]'] = {
'n': int(len(merged)),
'macro_f1': macro_f1(merged['true'], merged['pred']),
}
return out
def per_industry(g, preds, min_n=20):
out = {}
for ind, sub in g.groupby('industry'):
merged = sub.merge(preds.reset_index(), on='praxis_id', how='inner')
if len(merged) < min_n: continue
if len(set(merged['true'])) < 2: continue
out[ind] = {'n': int(len(merged)),
'macro_f1': macro_f1(merged['true'], merged['pred'])}
return out
def main():
print("=" * 70); print("CP-10: Slice evaluation"); print("=" * 70)
g = load_test()
preds_dict = find_preds()
print(f"Loaded gold-test {len(g)} rows; preds for {len(preds_dict)} models")
results = {}
for model_name, runs in preds_dict.items():
# Average per-row across seeds
if len(runs) == 1:
df = runs[0][1]
df['true'] = df['is_suggestion'].astype(str).str.lower().isin(['true','1']).astype(int) \
if 'is_suggestion' in df.columns else df.get('label', 0)
else:
# Combine seeds: majority vote per row by praxis_id
seed_dfs = [r[1] for r in runs]
base = seed_dfs[0][['praxis_id', 'is_suggestion']].copy()
for i, sdf in enumerate(seed_dfs):
base[f'pred_s{i}'] = sdf.set_index('praxis_id').loc[base['praxis_id']]['pred'].values
base['pred'] = base[[f'pred_s{i}' for i in range(len(seed_dfs))]].mean(axis=1).round().astype(int)
base['true'] = base['is_suggestion'].astype(str).str.lower().isin(['true','1']).astype(int)
df = base[['praxis_id', 'true', 'pred']]
# need pred + praxis_id + true for downstream
if 'praxis_id' not in df.columns:
print(f" skip {model_name}: no praxis_id")
continue
df = df[['praxis_id', 'pred']].set_index('praxis_id')
# Compute slices
merged_full = g.merge(df.reset_index(), on='praxis_id', how='inner')
if len(merged_full) == 0:
print(f" skip {model_name}: no overlap (test_n={len(g)})")
continue
results[model_name] = {
'n_overlap': int(len(merged_full)),
'overall': {
'macro_f1': macro_f1(merged_full['true'], merged_full['pred']),
'pos_f1': float(f1_score(merged_full['true'], merged_full['pred'], pos_label=1, zero_division=0)),
},
'per_form_recall': per_form(g, df),
'per_modifier_recall': per_modifier(g, df),
'per_specificity_recall': per_specificity(g, df),
'per_borderline_recall': per_borderline(g, df),
'per_macro': per_macro(g, df),
'per_length': per_length(g, df),
'per_industry': per_industry(g, df),
}
print(f" {model_name}: macro-F1 {results[model_name]['overall']['macro_f1']:.4f}, n_overlap={len(merged_full)}")
(RES / "slices.json").write_text(json.dumps(results, indent=2, default=str))
print(f"\nSaved {RES}/slices.json")
print("CP-10 DONE.")
if __name__ == '__main__':
main()
|