"""Main run: read hidden states while the model reads each text, told vs not told. Forward pass only, no generation. The text is the user turn of the chat template; the told condition prepends NOTE to that turn. Activations are pooled over the text's own tokens (the note's tokens are excluded) so the two conditions are comparable position for position. Saves mean-pooled and last-token vectors per layer. """ import json, pathlib, numpy as np, torch from transformers import AutoModelForCausalLM, AutoTokenizer from texts import TEXTS, NOTE, CONTROL_NOTE MODEL = "Qwen/Qwen2.5-3B-Instruct" OUT = pathlib.Path(__file__).parent / "results" OUT.mkdir(exist_ok=True) tok = AutoTokenizer.from_pretrained(MODEL) model = AutoModelForCausalLM.from_pretrained(MODEL, torch_dtype=torch.float16, device_map="cuda") model.eval() PREFIX = {"told": NOTE, "ctrl": CONTROL_NOTE, "plain": None} print("prefix tokens:", {k: len(tok(v)["input_ids"]) for k, v in PREFIX.items() if v}) def run(text: str, cond: str): user = (PREFIX[cond] + "\n\n" + text) if PREFIX[cond] else text prompt = tok.apply_chat_template([{"role": "user", "content": user}], tokenize=False, add_generation_prompt=False) start = prompt.index(text) enc = tok(prompt, return_offsets_mapping=True, return_tensors="pt") offs = enc.pop("offset_mapping")[0] span = [i for i, (a, b) in enumerate(offs.tolist()) if a >= start and b <= start + len(text)] with torch.no_grad(): hs = model(**enc.to("cuda"), output_hidden_states=True).hidden_states # (L+1) x [1, T, d] H = torch.stack([h[0] for h in hs]).float() # [L+1, T, d] return H[:, span].mean(1).cpu().numpy(), H[:, span[-1]].cpu().numpy(), len(span) meta, mean_told, mean_plain, mean_ctrl, last_told, last_plain, last_ctrl = [], [], [], [], [], [], [] for cat, texts in TEXTS.items(): for i, t in enumerate(texts): mt, lt, n = run(t, "told") mp, lp, _ = run(t, "plain") mc, lc, _ = run(t, "ctrl") mean_told.append(mt); last_told.append(lt); mean_plain.append(mp); last_plain.append(lp) mean_ctrl.append(mc); last_ctrl.append(lc) meta.append({"cat": cat, "i": i, "n_tokens": n, "text": t}) print(cat, i, n, flush=True) np.savez(OUT / "activations.npz", mean_told=np.stack(mean_told), mean_plain=np.stack(mean_plain), mean_ctrl=np.stack(mean_ctrl), last_told=np.stack(last_told), last_plain=np.stack(last_plain), last_ctrl=np.stack(last_ctrl)) (OUT / "meta.json").write_text(json.dumps(meta, ensure_ascii=False, indent=1), encoding="utf-8") print("saved", len(meta), "texts x 2 conditions x", mean_told[0].shape[0], "layers")