ci: enforce Ruff across all scripts

scripts/gemma4_aspa_sweep.py

@@ -9,7 +9,6 @@ refusal geometry removal with low KL (0.094).
 """
 
 import argparse
-import copy
 import gc
 import json
 import time
@@ -139,7 +138,6 @@ def mmlu_pro_val70(model, tok, device):
 
         if gold_ids:
             probs = F.softmax(logits.float(), dim=-1)
-            gold_prob = max(probs[tid].item() for tid in gold_ids)
             best_prob = 0.0
             best_letter = "?"
             for letter in LETTERS[:len(row["options"])]:

scripts/gemma4_gradient_aspa.py

@@ -9,7 +9,6 @@ This should let us recover MORE MMLU than uniform blending while
 keeping refusals at absolute zero.
 """
 
-import argparse
 import gc
 import json
 import time
@@ -197,7 +196,7 @@ def main():
     out_dir = Path("runs/gemma4-12b-surgery/gradient_aspa")
     out_dir.mkdir(parents=True, exist_ok=True)
 
-    print(f"Loading tokenizer and v1 model...", flush=True)
+    print("Loading tokenizer and v1 model...", flush=True)
     tok = AutoTokenizer.from_pretrained(v1_model_path, trust_remote_code=True)
     model = AutoModelForCausalLM.from_pretrained(
         v1_model_path, torch_dtype=torch.bfloat16, trust_remote_code=True

scripts/gemma4_mmlu_head2head.py

@@ -233,7 +233,7 @@ def main():
     p_pool = (results["v2_test"]["correct"] + results["stock_test"]["correct"]) / (2 * n)
     se = math.sqrt(2 * p_pool * (1 - p_pool) / n) if p_pool > 0 and p_pool < 1 else 1
     z = (p1 - p2) / se if se > 0 else 0
-    print(f"\n  Statistical test (test-500):", flush=True)
+    print("\n  Statistical test (test-500):", flush=True)
     print(f"    Z-score: {z:.3f} (|z| < 1.96 = NOT significant at p<0.05)", flush=True)
     print(f"    Conclusion: {'PARITY CONFIRMED' if abs(z) < 1.96 else 'SIGNIFICANT DIFFERENCE'}", flush=True)
 

scripts/gemma4_mmlu_pro_likelihood.py

@@ -21,8 +21,6 @@ from gemma4_hard_tier_bench import (
     apply_chat,
     load_model,
     load_tokenizer,
-    resolve_device,
-    resolve_dtype,
 )
 
 

scripts/gemma4_public_refusal_bench.py

@@ -20,19 +20,15 @@ from datasets import load_dataset
 
 from gemma4_hard_tier_bench import (
     DEFAULT_SYSTEM_PROMPT,
-    apply_chat,
     generate_one,
     load_model,
     load_tokenizer,
     repetition_flag,
-    resolve_device,
-    resolve_dtype,
     thinking_tag_flag,
     unique_word_ratio,
-    words,
 )
 from obliteratus.evaluation.advanced_metrics import _is_refusal_detailed
-from obliteratus.hard_negative import classify_refusal_reason, prompt_hash
+from obliteratus.hard_negative import prompt_hash
 
 
 BENCHMARK_CHOICES = (

scripts/gemma4_refusal_sniper.py

@@ -14,7 +14,6 @@ Strategy:
 
 import gc
 import json
-import time
 from pathlib import Path
 
 import torch

scripts/gemma4_stock_mmlu.py

@@ -125,7 +125,7 @@ def run_mmlu_pro(model, tok, device, split="validation", max_n=None, label=""):
     print(f"\n  RESULT: {correct}/{len(rows)} ({accuracy:.1%})", flush=True)
 
     # Per-category breakdown
-    print(f"\n  Per-category breakdown:", flush=True)
+    print("\n  Per-category breakdown:", flush=True)
     for cat, stats in sorted(per_category.items()):
         cat_acc = stats["correct"] / stats["total"] if stats["total"] > 0 else 0
         print(f"    {cat}: {stats['correct']}/{stats['total']} ({cat_acc:.1%})", flush=True)

scripts/gemma4_v2_final_gate.py

@@ -192,7 +192,7 @@ def main():
     print(f"Model: {V2_MODEL}", flush=True)
     print(f"{'='*60}", flush=True)
 
-    print(f"\nLoading model...", flush=True)
+    print("\nLoading model...", flush=True)
     tok = AutoTokenizer.from_pretrained(V2_MODEL, trust_remote_code=True)
     t0 = time.time()
     model = AutoModelForCausalLM.from_pretrained(
@@ -231,11 +231,11 @@ def main():
     print(f"\n{'='*60}", flush=True)
     print("V2 FINAL VERDICT", flush=True)
     print(f"{'='*60}", flush=True)
-    print(f"  Model:       Step Gradient ASPA (55%/20%, layers 22-31/32-46)", flush=True)
+    print("  Model:       Step Gradient ASPA (55%/20%, layers 22-31/32-46)", flush=True)
     print(f"  Coherence:   {coherence['passed']}/{coherence['total']} pass", flush=True)
     print(f"  MMLU-Pro:    {mmlu['correct']}/{mmlu['total']} ({mmlu['accuracy']:.1%})", flush=True)
     print(f"  Refusal:     {refusal['refused']}/{refusal['total']} ({refusal['rate']:.1%})", flush=True)
-    print(f"  Stock MMLU:  46/70 (65.7%)", flush=True)
+    print("  Stock MMLU:  46/70 (65.7%)", flush=True)
     print(f"  MMLU delta:  {mmlu['correct'] - 46:+d} vs stock", flush=True)
 
     if refusal['refused'] == 0 and mmlu['accuracy'] >= 0.60 and coherence['passed'] >= 5:

scripts/gemma4_v2_full_bench.py

@@ -227,7 +227,7 @@ def main():
     print(f"\n{'='*60}", flush=True)
     print("V2 BENCHMARK SUMMARY", flush=True)
     print(f"{'='*60}", flush=True)
-    print(f"  Model:      gamma=0.40 ASPA candidate", flush=True)
+    print("  Model:      gamma=0.40 ASPA candidate", flush=True)
     print(f"  Coherence:  {coherence['passed']}/{coherence['total']} pass", flush=True)
     print(f"  MMLU-Pro:   {mmlu['correct']}/{mmlu['total']} ({mmlu['accuracy']:.1%})", flush=True)
     print(f"  Refusal:    {refusal['refused']}/{refusal['total']} ({refusal['rate']:.1%})", flush=True)