Upload densesteer_repro.py with huggingface_hub

densesteer_repro.py

@@ -425,8 +425,34 @@ def stage_nll(args):
     (OUT/"nll_summary.json").write_text(json.dumps(rows, indent=2))
     log("wrote nll.csv")
 
+def stage_accsweep(args):
+    """Reproduce Fig 5: GSM8K accuracy vs steering coefficient lambda at L17
+    (and optionally other layers), including lambda=0 baseline, on a fixed subset."""
+    vecs = torch.load(OUT/"steering_vectors.pt")
+    model, tok = load_model(TARGET)
+    g = load_gsm8k("test")
+    n = args.acc_n
+    items = [{"question": g[i]["question"], "gold": gsm8k_gold(g[i]["answer"])} for i in range(n)]
+    gp = lambda it: chat(tok, COT_PROMPT.format(problem=it["question"]))
+    lams = [float(x) for x in os.environ.get("DS_ACC_LAMS", "-6,-2,0,1,2,3,4,6,8").split(",")]
+    layers = [int(x) for x in os.environ.get("DS_ACC_LAYERS", str(STEER_LAYER)).split(",")]
+    rows = []
+    for L in layers:
+        for lam in lams:
+            steer = None if lam == 0 else (L, vecs[L], lam)
+            acc, _ = _eval_set(model, tok, items, gp, gsm8k_correct, "gold", MAX_NEW, args.bs, steer=steer)
+            rows.append({"layer": L, "lambda": lam, "gsm8k_acc": acc, "n": n})
+            log(f"accsweep L{L} lam={lam} acc={acc:.1f}")
+    import csv
+    with open(OUT/"accsweep.csv", "w", newline="") as f:
+        w = csv.DictWriter(f, fieldnames=["layer","lambda","gsm8k_acc","n"]); w.writeheader(); w.writerows(rows)
+    (OUT/"accsweep.json").write_text(json.dumps(rows, indent=2))
+    log("wrote accsweep.csv")
+    del model; torch.cuda.empty_cache()
+
 STAGES = {"calib": stage_calib, "rewrite": stage_rewrite, "steervec": stage_steervec,
-          "eval": stage_eval, "layersweep": stage_layersweep, "nll": stage_nll}
+          "eval": stage_eval, "layersweep": stage_layersweep, "nll": stage_nll,
+          "accsweep": stage_accsweep}
 
 def main():
     ap = argparse.ArgumentParser()
@@ -438,6 +464,7 @@ def main():
     ap.add_argument("--limit", type=int, default=0)
     ap.add_argument("--sweep_n", type=int, default=100)
     ap.add_argument("--nll_n", type=int, default=200)
+    ap.add_argument("--acc_n", type=int, default=300)
     ap.add_argument("--bs", type=int, default=16)
     args = ap.parse_args()
     log("stages:", args.stages, "| out:", OUT.resolve())