J-lens workspace reproduction + loop retrofit: lens, band looping, adapters, controls, multi-task evals

Reproduction of the 2026 workspace/J-lens paper on gemma-4 (E2B/12B/26B),
plus the workspace-loop retrofit line: merge adapter, prompt-only latent
planning (MBPP), carry variant, attribution controls (FF/pause/untrained),
band-location ablation, Blocksworld harness, 12B replication scripts.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

scripts/train_carry.py

@@ -0,0 +1,149 @@
+"""Stage B / design C: train the merge adapter for prompt-prefill + carry.
+
+GSM8K only (the task the prompt-only loop failed on). Sequence:
+[prompt] [p x <unused0> pauses] [gold answer]; k=2 prefill loops on the
+prompt; carry scan through pauses + answer; CE on answer tokens.
+Curriculum: easy p=2, hard p=6 (hard needs the longer latent chain).
+"""
+
+import argparse
+import json
+import math
+import random
+import sys
+import time
+from pathlib import Path
+
+import torch
+import torch.nn.functional as F
+
+from carry_common import PAUSE_ID, carry_logits
+from loop_common import BandLooper, MergeAdapter, chat_prompt, DIRECT_SUFFIX
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+from jlens.core import load_model  # noqa: E402
+
+OUT = Path(__file__).resolve().parent.parent / "results-loop"
+STEPS = 600
+BATCH = 4
+LR = 1e-3
+WARMUP = 20
+K_PREFILL = 2
+P_BY_LABEL = {"easy": 2, "hard": 6}
+
+ap = argparse.ArgumentParser()
+ap.add_argument("--feedforward", action="store_true",
+                help="pause-token control: adapter(e,e), no carry")
+ap.add_argument("--seed", type=int, default=0)
+ARGS = ap.parse_args()
+SEED = ARGS.seed
+TAG = "pausectl" if ARGS.feedforward else "carry"
+
+
+def lr_at(step):
+    if step < WARMUP:
+        return LR * (step + 1) / WARMUP
+    t = (step - WARMUP) / max(1, STEPS - WARMUP)
+    return 1e-4 + 0.5 * (LR - 1e-4) * (1 + math.cos(math.pi * t))
+
+
+def build_batch(tok, items, p, device="cuda"):
+    seqs, labs, plens = [], [], []
+    for it in items:
+        pr = tok(chat_prompt(tok, it["question"], DIRECT_SUFFIX),
+                 add_special_tokens=False)["input_ids"]
+        a = tok(it["gold"] + "<end_of_turn>",
+                add_special_tokens=False)["input_ids"]
+        seqs.append(pr + [PAUSE_ID] * p + a)
+        labs.append([-100] * (len(pr) + p) + a)
+        plens.append(len(pr))
+    T = max(len(s) for s in seqs)
+    pad = tok.pad_token_id or 0
+    ids = torch.full((len(seqs), T), pad, dtype=torch.long)
+    lab = torch.full((len(seqs), T), -100, dtype=torch.long)
+    msk = torch.zeros((len(seqs), T), dtype=torch.long)
+    for i, (s, l) in enumerate(zip(seqs, labs)):
+        ids[i, : len(s)] = torch.tensor(s)
+        lab[i, : len(s)] = torch.tensor(l)
+        msk[i, : len(s)] = 1
+    return (ids.to(device), msk.to(device), lab.to(device),
+            torch.tensor(plens, device=device))
+
+
+@torch.no_grad()
+def val_loss(looper, adapter, tok, items, p, k=K_PREFILL):
+    tot, n = 0.0, 0
+    for i in range(0, len(items), BATCH):
+        ids, msk, lab, plens = build_batch(tok, items[i : i + BATCH], p)
+        logits = carry_logits(looper, adapter, ids, msk, plens, k, feedforward=ARGS.feedforward)
+        loss = F.cross_entropy(logits[:, :-1].flatten(0, 1).float(),
+                               lab[:, 1:].flatten(), ignore_index=-100)
+        tot += loss.item() * len(ids)
+        n += len(ids)
+    return tot / n
+
+
+def main():
+    rng = random.Random(SEED)
+    torch.manual_seed(SEED)
+    data = [it for it in json.load(open(OUT / "star_data.json"))
+            if it["split"] == "train" and it["label"] != "drop"]
+
+    model, tok = load_model(dtype=torch.bfloat16)
+    for pp in model.parameters():
+        pp.requires_grad_(False)
+    looper = BandLooper(model)
+    adapter = MergeAdapter().cuda()
+    opt = torch.optim.AdamW(adapter.parameters(), lr=LR, weight_decay=0.01)
+
+    keep = [it for it in data
+            if len(tok(it["question"])["input_ids"]) + 30 <= 400]
+    rng.shuffle(keep)
+    pool = {l: [it for it in keep if it["label"] == l]
+            for l in ("easy", "hard")}
+    val = {l: pool[l][:16] for l in pool}
+    pool = {l: pool[l][16:] for l in pool}
+    print(f"pool: easy={len(pool['easy'])} hard={len(pool['hard'])}",
+          flush=True)
+
+    log = []
+    t0 = time.time()
+    for step in range(STEPS):
+        lbl = ("easy", "hard")[step % 2]
+        batch = rng.sample(pool[lbl], BATCH)
+        p = P_BY_LABEL[lbl]
+        ids, msk, lab, plens = build_batch(tok, batch, p)
+
+        for g in opt.param_groups:
+            g["lr"] = lr_at(step)
+        logits = carry_logits(looper, adapter, ids, msk, plens, K_PREFILL, feedforward=ARGS.feedforward,
+                              use_checkpoint=True)
+        loss = F.cross_entropy(logits[:, :-1].flatten(0, 1).float(),
+                               lab[:, 1:].flatten(), ignore_index=-100)
+        opt.zero_grad(set_to_none=True)
+        loss.backward()
+        torch.nn.utils.clip_grad_norm_(adapter.parameters(), 1.0)
+        opt.step()
+
+        log.append({"step": step, "label": lbl, "p": p, "loss": loss.item()})
+        if step % 10 == 0:
+            print(f"step {step:4d}  {lbl} p={p}  loss={loss.item():.4f}  "
+                  f"({(time.time()-t0)/(step+1):.1f}s/step)", flush=True)
+        if step % 100 == 99 or step == STEPS - 1:
+            vals = {}
+            for l in ("easy", "hard"):
+                for pp_ in (0, 2, 6):
+                    vals[f"{l}_p{pp_}"] = val_loss(looper, adapter, tok,
+                                                   val[l], pp_)
+            print(f"  val@{step}: " + "  ".join(
+                f"{n}={v:.3f}" for n, v in sorted(vals.items())), flush=True)
+            log.append({"step": step, "val": vals})
+            torch.save(adapter.state_dict(),
+                       OUT / f"adapter_{TAG}_e{step+1}.pt")
+            json.dump(log, open(OUT / f"train_{TAG}_log.json", "w"), indent=1)
+
+    print("done", flush=True)
+
+
+if __name__ == "__main__":
+    main()