tlab-looped-transformer

Weight-tied looped transformer: one 3-layer Qwen3-style block applied r times, 448-dim, 9,064,608 parameters, trained from scratch on FineWeb next-token prediction. T-Lab test task submission. Run full_control90_kaggle, 90.0M tokens, step 43944.

Results

metric	value
CE @ 1 loop	3.9622
best val CE	3.6599 (at 10 loops)
val perplexity	38.86
bits/byte	1.5829 (at 3.3358 bytes/token)
useful-depth plateau	[6, 17] on the dense 1..64 eval grid
loop gain (CE@1 − CE@best)	0.3023

Perplexity is tokenizer-dependent and this model uses its own 4096-token BPE, so it is not comparable across submissions; bits/byte is the figure that survives a change of tokenizer.

These weights are the DENSE control, not the recipe the report recommends

METHOD.md §2 recommends supervision annealing — dense supervision for most of training, then terminal-loop-only for the final ~10% of steps. These weights do not implement it. They are the dense control: supervise_k = 5 throughout, no supervise_k_final, no supervise_switch_frac.

That is deliberate. The annealing recommendation is about where depth stays useful — the useful band widens at 6 of 6 seeds — and not about the loss: its CE half was withdrawn at n = 4, and the six-point mean is −0.0247, inside the replicate floor. So the recipe that buys depth and the checkpoint that gives the lowest loss are different arms, and the lowest-loss one ships. A 10M-token checkpoint of the annealed recipe exists (rec_sw90_s2); no 90M one does. See METHOD.md §4.

Why this checkpoint, when the report contains a better perplexity

A reader who finds 37.52 in the report and 38.86 here should know this was a decision, not an error. A norm-penalty arm at the same 90M budget reaches 37.52 perplexity. It is not what ships, for four reasons measured rather than asserted:

1. 88% of its apparent loop-gain advantage is loop-1 damage (ΔCE@1 = +0.2263). It wins the loop-gain statistic by making one loop worse, not by making depth worth more.
2. Its useful band narrows, [6,17] → [6,14]. On the axis this task actually asks about — value from many loops — it is the worse model.
3. It is the only arm whose map converges (ρ = 0.9953 / 0.9915 at loops 32/64), which is the regime the report's §2 argues against on independent evidence.
4. It carries a clipping confound the stored artifacts cannot resolve.

The control has no confound on either axis. See report.md §4.6b and §6.0b/D3, and submission/METHOD.md §4.

Counting the parameters — the obvious way gives the wrong answer

sum(v.numel() for v in state_dict.values()) returns 10,899,616, which is over the task's 10M cap. That is an artifact of weight tying, this architecture's central feature: lm_head and embed are the same nn.Parameter under two names, so a state_dict sum counts the tied embedding twice. The difference is exactly vocab x hidden = 1,835,008.

sum(p.numel() for p in m.parameters())   # 9,064,608  <- the real count, and what every number uses

Files, and why the tokenizer is one of them

• model.pt — weights (torch.load, weights_only=False; contains model, model_cfg, train_cfg)
• tokenizer.json — the vocabulary these weights were trained with. Do not substitute another one and do not retrain it: a mismatch raises nothing and reports CE ≈ ln(4096) = 8.3178, i.e. chance, which looks like a broken model rather than a broken setup.
• model.py — the architecture, so this checkpoint loads without cloning the GitHub repo.

Verify the download before trusting a number

python src/check_tokenizer_identity.py <this checkpoint> --expect-ce1 3.9622

That gate judges vocabulary against chance and protocol drift against the sample's own SEM, so it distinguishes "wrong vocabulary" from "slightly different eval batch". Expect |diff| well under 0.1.

import torch
from model import Config, LoopedTransformer
ck = torch.load("model.pt", map_location="cpu", weights_only=False)
m = LoopedTransformer(Config(**ck["model_cfg"])); m.load_state_dict(ck["model"]); m.eval()
logits_per_loop, state_norms = m(input_ids, n_loops=10, return_all_loops=True)

This checkpoint's own state norms

loop	1	8	16	64
‖h‖	466.6	2334.4	3977.3	12424.4

Stated because the released models differ in state scale by up to 380x. Any absolute clamp/threshold level in the report was measured on one checkpoint and does not transfer; derive levels from these numbers or re-run src/radial_clamp.py on this checkpoint, which does it for you.

Config

{"vocab_size": 4096, "hidden_size": 448, "n_heads": 4, "n_kv_heads": 2, "head_dim": 112, "intermediate_size": 1344, "layers_per_loop": 3, "n_prelude": 0, "n_coda": 0, "rms_norm_eps": 1e-06, "rope_theta": 10000.0, "max_position_embeddings": 512, "readout_mode": "norm", "convex_gate": false, "explore_noise": 0.0, "explore_anneal": true, "fixed_gate": null, "truncate_bptt": null, "state_renorm": false, "inject_mode": "additive", "depth_init": true, "residual_scale": null, "scale_clock": false, "gate_alpha_init": 0.874, "n_loop_eff": 24, "cond_mode": "none", "cond_lora_rank": 4, "cond_lora_branches": 4, "cond_fixed_branch": null, "kv_untie_buckets": 1, "depth_gate_mode": "none", "xsa": false, "kv_window": 1}

See the GitHub repo's report.md for the full ablation set, the negative results, and the failure log (§6.0).