Add dataset card

README.md

@@ -1,149 +1,24 @@
-# LoopBench
-
-A length-generalization study of **looped transformers** on algorithmic tasks,
-run entirely on a single Apple Silicon machine (M4 Max, PyTorch/MPS). Tiny
-decoder-only models (~0.8–3M params) are trained on seeded algorithmic
-generators and evaluated for out-of-distribution (OOD) length extrapolation.
-
-The confirmed R=8 result: **the depth-scaling exponent `p` improves held-out
-Dyck-1 length extrapolation without materially changing ID accuracy**. An
-independent, predeclared 15-pair replication improved far-OOD exact match from
-.403 to .614 (+.211, paired t=2.734, 95% CI [.046, .377]); its class-balanced
-sensitivity check also replicated. Addition remains at 0% OOD and parity
-returns to chance despite better ID training. The completed five-task R=8 sweep
-also finds that p=.5 strongly improves prefix-sum ID trainability (.649→.953;
-exploratory paired t=2.816) and removes a copy seed failure, but both tasks stay
-at effectively 0% OOD: better optimization is not algorithmic extrapolation.
-An exploratory inference-depth probe finds that the Dyck advantage is already
-present at R=1 and peaks at R=4, so it does not support a simple mechanism where
-`p` merely prevents degradation in later loop iterations.
-
-## Method in one screen
-
-- **Model** (`loopbench/model.py`): decoder-only, RMSNorm, RoPE (base 10000),
-  SwiGLU. Two block styles:
-  - `preln` — standard pre-norm residual block (the control).
-  - `loopdeepnorm` — [DeepLoop's](https://arxiv.org/abs/2607.13491) hybrid
-    sandwich, `α = (2N)^p` on the residual, `β = (8N)^(−p)` applied once at
-    init to the output projections, `N = K · R_ref`. `p = 0` ⇒ `α = β = 1`
-    (unscaled).
-- **Loop**: `K` physical blocks applied `R` times; optional Bansal-style input
-  injection (`recall = add`). Truncated backprop keeps grad for the last
-  `backprop_window = 8` iterations. This gives full gradient coverage at R=8;
-  fixed k=8 is **not** comparable at larger R.
-- **Schedules** (`loopbench/schedules.py`): `fixed`, `uniform`,
-  `poisson_lognormal`; `n_ref ∈ {max, mean, per_batch}` sets the `N` behind
-  `α/β` (research question RQ2).
-- **Tasks** (`loopbench/tasks/`): binary addition, parity, copy, Dyck-1,
-  prefix-sums (mod 7). Pure-Python, seeded, emitting `[runs, batch, seq]`
-  tensors directly. **Full-answer prediction (FAP)**: the model reads
-  `<question> EOQ` and predicts the whole `<answer> EOS`; loss is masked to the
-  answer span; the metric is exact match over that span.
-
-## Throughput levers (built in from day one)
-
-Tiny models on MPS are overhead-bound, so the grid's wall-clock is attacked by:
-
-1. **Run-stacking** (`loopbench/stacking.py`) — all seeds of a cell train as one
-   `vmap`-ed model (`stack_module_state` + `functional_call` + `vmap`). One
-   process, N runs, ~one run's wall time. bf16 is applied by explicit casting,
-   not `torch.autocast` (which does not compose under `vmap` on MPS).
-   Memory-heavy copy/prefix cells use deterministic seed shards `(0,1)`, `(2,3)`,
-   `(4)` and merge their logs; this changes parallelism, not experiment factors.
-2. **No hidden syncs** — metrics accumulate on-device; host transfer every
-   `log_every` steps; divergence is a device-side check.
-3. **Truncated backprop** through the loop (in the model).
-4. **Early stopping** with a token ledger.
-5. **On-device pregenerated data** (`loopbench/data.py`) — one pool tensor,
-   moved to the device once, indexed per step. No DataLoader.
-
-Per-run divergence is masked out so one diverging seed cannot poison its
-stack-mates.
-
-## Quickstart
-
-```bash
-uv sync --frozen
-uv run pytest                     # unit tests (tasks, model, stacking, schedules)
-
-# Stage 0 bring-up + divergence probe (the gate before the grid)
-uv run python scripts/stage0.py
-
-# A single stacked job
-uv run python -m loopbench.train --task addition --seeds 0,1,2,3,4 \
-    --k-layers 2 --R 8 --recall add --device mps
-
-# Rebuild the committed Stage-1a summary and hero figure
-uv run python scripts/analyze_stage1a.py
-
-# Rebuild the exploratory Dyck inference-depth probe
-uv run python scripts/analyze_iteration_dynamics.py
-```
-
-Outputs: per-job JSON in `runs/<name>.json` (config, per-seed training curves,
-ID/OOD exact-match by eval length × inference-R, per-iteration activation norms,
-per-block grad norms, divergence flags, token ledger, tokens/sec) and per-seed
-checkpoints in `checkpoints/<name>/seed_<s>.pt` for full reevaluation.
-The committed CSV/JSON/SVG evidence can be regenerated without the raw training
-runs. Full reevaluation requires the separately distributed checkpoint bundle.
-
-## Comparability pins (read this before comparing numbers)
-
-Results here are **internally comparable at R=8 and only qualitatively comparable
-to prior papers**. The pins for the confirmed R=8 study are:
-
-- **Tokenizer**: task-specific symbol vocabularies (< 50 symbols), shared
-  special tokens (`PAD=0`, `EOS=1`, `EOQ=2`).
-- **Supervision**: full-answer prediction (FAP), matching Fan et al. Next-token
-  variants are out of scope.
-- **Positions**: RoPE, base 10000 (NoPE available as a flag but PE ablations are
-  out of scope).
-- **Optimizer**: Muon on attention/MLP matrices (muon_lr 0.01) + AdamW
-  β=(0.9, 0.95), lr 3e-4 on embeddings/head/norms; cosine schedule, warmup, and
-  grad clip 1.0. These LoopDeepNorm values were frozen for Stage 1a. Embeddings
-  use init std 0.02 to remove the cold-start logit spike.
-- **Precision**: bf16 on MPS (fp32 fallback off-device). Seeds control
-  data + init; run-level numeric noise on MPS is reported as part of the
-  seed variance.
-- **Checkpoint/evaluation**: weights are selected only by ID validation seed
-  4242, then reloaded and evaluated on shared held-out seed 20260721.
-- **Causal scope**: `p` changes residual scaling and output-projection
-  initialization together. Without a non-recurrent model trained under the same
-  α/β values, the Dyck result is not evidence that the benefit is specific to
-  weight tying or recurrence.
-- **Gradient coverage**: k=8 covers all R=8 iterations. The aborted R=32 gate had
-  only 25% coverage, so it is documented as a blocked method extension rather
-  than evidence about high-recurrence capacity.
-
-## What this contributes
-
-- A small Apple-Silicon research harness for stacked independent runs,
-  per-seed early stopping, best-checkpoint capture, and memory-safe evaluation.
-- A reproducible separation between improved ID optimization and genuine length
-  extrapolation across five tasks, with checkpoint hashes and generated evidence
-  artifacts. Prefix-sum is the sharp negative control: large ID gain, zero OOD.
-- A methodological warning for recurrent-depth studies: fixed truncated-BPTT
-  windows change credit coverage as R changes.
-- A bounded practical result: a separate one-seed HDFS short-to-long anomaly
-  gate did not validate a loop-model advantage; the matched Transformer won on
-  far-OOD AUPRC (.699 versus .509), so the transfer study stopped at its gate.
-
-Stage 1a is consolidated in [STAGE1A_RESULTS.md](STAGE1A_RESULTS.md). The R=32
-extension is deferred until k/R gradient comparability is solved.
-
-## Documentation
-
-- [STAGE1A_RESULTS.md](STAGE1A_RESULTS.md) — final metrics, statistical tests,
-  limitations, and evidence inventory.
-- [DYCK_REPLICATION_PLAN.md](DYCK_REPLICATION_PLAN.md) — frozen independent
-  replication protocol, committed before seeds 15–29 were trained.
-- [RESEARCH_NOTES.md](RESEARCH_NOTES.md) — RASP-L framing and practical-transfer
-  rationale.
-- [STAGE1_PLAN.md](STAGE1_PLAN.md) — frozen experiment contract.
-
-## Citation / license / contact
-
-Citation metadata is in [CITATION.cff](CITATION.cff).
-LoopBench is available under the [Apache License 2.0](LICENSE).
-
-Vadim Smirnov · `ukint-vs@proton.me`
+---
+license: apache-2.0
+pretty_name: LoopBench Checkpoints and Evidence
+tags:
+  - transformers
+  - length-generalization
+  - looped-transformers
+---
+
+# LoopBench checkpoints and evidence
+
+This repository contains the separately distributed artifacts for
+[LoopBench](https://github.com/ukint-vs/loopbench) v0.1.0.
+
+- `checkpoints/` — 125 selected PyTorch checkpoints.
+- `runs/` — raw training JSON logs.
+- `artifacts/` — derived CSV/JSON evidence, checkpoint SHA-256 hashes, and the
+  summary figure.
+
+The code, experiment contracts, results, and reproduction commands live in the
+LoopBench GitHub repository. All 125 checkpoint hashes were verified against
+`artifacts/stage1a_metrics.csv` before publication.
+
+Licensed under Apache-2.0.