update docs

README.md

@@ -9,6 +9,7 @@ tags:
   - s23dr
   - cvpr-2026
 datasets:
+  - usm3d/hoho22k_2026_trainval
   - usm3d/s23dr-2026-sampled_4096_v2
   - usm3d/s23dr-2026-sampled_2048_v2
 metrics:
@@ -16,161 +17,178 @@ metrics:
 pipeline_tag: other
 ---
 
-# S23DR 2026 Learned Baseline
+# S23DR 2026 - Learned Submission
 
-A learned baseline for the **S23DR 2026** challenge (**S**tructured and **S**emantic **3D R**econstruction, or S^2 3DR), part of the [USM3D workshop](https://usm3d.github.io) at CVPR 2026. The model takes a fused point cloud of a building and predicts its wireframe as a set of 3D line segments.
+A learned-model baseline for the
+[S23DR 2026](https://huggingface.co/spaces/usm3d/S23DR2026) wireframe-estimation
+challenge. Counterpart to
+[`usm3d/handcrafted_submission_2026`](https://huggingface.co/usm3d/handcrafted_submission_2026):
+the handcrafted entry is rule-based; this one is a Perceiver transformer
+trained on fused 3D point clouds.
 
-**Headline result: HSS = 0.382** on the 1024-sample validation set (shipped checkpoint).
+## Task
 
-For context, the handcrafted baseline scores HSS = 0.307 on the same split.
+Per scene from
+[`usm3d/hoho22k_2026_trainval`](https://huggingface.co/datasets/usm3d/hoho22k_2026_trainval):
+multi-view RGB, ADE20K and Gestalt segmaps, MoGe metric depth, COLMAP SfM
+points and camera poses (BPO + COLMAP). Predict a 3D building wireframe -
+a list of vertices and the edges connecting them. Scored by [HSS](https://arxiv.org/abs/2503.08208) on a held-out
+test split.
 
-## Quick start
+## Run inference
 
-Run the submission pipeline directly (matches the competition eval harness):
-
-```bash
+```
 python script.py
 ```
 
-That loads `checkpoint.pt`, fuses the input views into a 4096-point cloud, runs the model, and writes the predicted wireframe for each scene.
+The challenge harness provides `params.json` (with the test-dataset id) and
+runs `script.py`. The script downloads the dataset to `/tmp/data`, loads
+`checkpoint.pt`, iterates over the validation + test splits, and writes
+`submission.json` - one entry per sample of the form
+`{order_id, wf_vertices, wf_edges}`.
 
-To reproduce the checkpoint from scratch on a single RTX 4090 (~3 hours):
+Requires a CUDA GPU plus `torch`, `huggingface_hub`, `datasets`, `numpy`,
+`opencv-python`, `scipy`, and `tqdm`.
 
-```bash
-bash reproduce.sh
-```
+## Pipeline
 
-Or for a bit-identical deterministic run (~5.5 hours, slower because it disables `torch.compile`):
-
-```bash
-bash reproduce_deterministic.sh
+```
+raw multi-view sample
+  -> point fusion          per-view depth unprojection + COLMAP, labeled
+                           by ADE/Gestalt              (point_fusion.py)
+  -> priority sampling     subsample to 4096 pts: 3072 COLMAP + 1024 depth
+                                                       (make_sampled_cache.py)
+  -> Perceiver             hidden=256, 256 latents x 7 layers, 64 segment
+                           queries; each query -> (midpoint, direction,
+                           length) + confidence        (model.py)
+  -> postprocess           conf > 0.5 -> segments -> iterative vertex
+                           merge -> snap to point cloud -> horizontal snap
+                                                       (segment_postprocess.py,
+                                                        postprocess_v2.py)
+  -> {order_id, wf_vertices, wf_edges}                 -> submission.json
 ```
 
-Both scripts run the full three-stage recipe described below. See `REPRODUCE.md` for the exact hyperparameters and the reproducibility notes.
+## Reproduce training
 
-## Architecture
+See [REPRODUCE.md](REPRODUCE.md) for the full recipe, ablations, and
+reproduction notes. Three stages, ~3hr on a single RTX 4090. All HSS numbers
+in this section are on **dev val** (the last 1024 scenes of the published
+training set; see "Evaluation sets" in REPRODUCE.md). The per-stage column
+below is from the **original training run**, not from re-running the
+script - the original run is gone, and the published recipe does not exactly
+reproduce it.
 
-A Perceiver-style transformer that ingests the point cloud as a sequence of per-point tokens and decodes a fixed set of 3D line segments through cross-attention into a latent.
+| Stage                            | Steps | LR    | Batch | Dev val HSS (orig) |
+|----------------------------------|------:|-------|------:|--------------------|
+| 2048 from scratch                |  125k | 3e-4  |    32 | 0.281              |
+| 4096 finetune                    |   10k | 3e-5  |    64 | 0.351              |
+| Cooldown + endpoint loss         |   35k | 3e-5  |    64 | **0.382**          |
 
 ```
-Perceiver: hidden=256, ff=1024
-  latent_tokens=256, latent_layers=7
-  encoder_layers=4, decoder_layers=3, cross_attn_interval=4
-  num_heads=4, kv_heads_cross=2, kv_heads_self=2
-  qk_norm=L2, rms_norm=True, dropout=0.1
-  segments=64, segment_param=midpoint_dir_len, segment_conf=True
-  behind_emb_dim=8, vote_features=True, activation=gelu
+bash reproduce.sh                # ~3hr,   torch.compile,  dev val HSS typ. 0.375-0.379
+bash reproduce_deterministic.sh  # ~5.5hr, no compile,     bit-reproducible at 0.372
+bash make_datasets.sh            # rebuild sampled datasets from raw tars
 ```
 
-The decoder predicts 64 candidate segments, each parametrized as midpoint + direction + length with a confidence head. Training uses a Sinkhorn optimal-transport loss to match predicted segments to ground-truth, plus a symmetric endpoint L1 term in the cooldown stage.
-
-All architecture and optimizer settings live in `configs/base.json`.
-
-## Training recipe
-
-The model ships with a three-stage recipe. Each stage starts from the previous stage's final checkpoint.
-
-| Stage | Input | Steps | LR | Batch | Notes | HSS |
-|---|---|---|---|---|---|---|
-| 1. 2048 from scratch | 2048 pts | 0 -> 125k | 3e-4, warmup 10k | 32 | Random init, sinkhorn only | 0.281 |
-| 2. 4096 finetune | 4096 pts | 125k -> 135k | 3e-5 constant | 64 | Gentle LR preserves representations | 0.351 |
-| 3. Endpoint cooldown | 4096 pts | 135k -> 170k | 3e-5 then linear decay | 64 | Adds endpoint L1 loss, tightens vertices | **0.382** |
-
-**Why 2048 first:** training directly on 4096 overfits (1.47x train/val ratio vs 1.19x for 2048). Starting on 2048 produces better-generalized representations that the 4096 finetune can then specialize.
-
-**Why a gentle LR on finetune:** LR > 1e-4 causes catastrophic forgetting of the 2048 geometry understanding.
-
-**Why endpoint loss only in stage 3:** the Sinkhorn loss operates on the midpoint/direction/length parametrization and doesn't directly penalize vertex position error. Adding a symmetric endpoint L1 against the detached Sinkhorn assignment tightens vertex precision in the cooldown.
-
-Full details, including the "what does not work" list (BuildingWorld pretraining, mixed training, high dropout, etc.), are in `REPRODUCE.md`.
-
-## Evaluation
-
-> **About the numbers:** all val scores below are HSS at confidence threshold 0.7, averaged over the 1024-sample *internal validation split* we hold out from the published training data (`usm3d/s23dr-2026-sampled_{2048,4096}_v2:validation`). They are **not** test-set numbers. The only test-set number we have is the public leaderboard score of the older 2048 submission (see `submitted_2048/` and the last table below).
->
-> All numbers below are freshly measured in this release against the checkpoints in this repo.
-
-### Shipped model and reproductions
-
-| Model | Checkpoint | HSS @ 4096 | HSS @ 2048 |
-|---|---|---|---|
-| Handcrafted baseline | — | 0.307 | — |
-| **Current release (shipped)** | `checkpoint.pt` | **0.3819** | 0.3734 |
-| Closest compiled E2E repro (#4) | `repro_runs/e2e_repro4_hss379/` | 0.3736 | 0.3675 |
-| Best compiled repro from this codebase | `repro_runs/compiled_repro_hss376/` | 0.3757 | 0.3670 |
-| Deterministic E2E repro (bit-reproducible) | `repro_runs/deterministic_hss372/` | 0.3716 | 0.3665 |
-
-All repros use the exact 3-stage recipe on a single RTX 4090. The shipped `checkpoint.pt` was trained on the same recipe before this release branch was cut; the ~0.005-0.010 HSS gap between shipped and repros is compiled-mode run-to-run variance (see the Reproducibility section).
+Documented compiled E2E runs from this codebase fall in **dev val HSS
+0.342-0.379**, with the modal cluster at 0.375-0.379 and low-side excursions
+at 0.349 and 0.342; no run reaches the original 0.382. Best compiled repro is
+**0.376** (`repro_runs/compiled_repro_hss376/`). The deterministic recipe is
+bit-identical across runs at 0.372 but follows a different numerical
+trajectory than compiled mode (it is not a reproduction of any compiled run).
 
-### Training progression (deterministic repro, all stages measured fresh)
-
-| Stage | Steps | HSS @ 4096 | HSS @ 2048 |
-|---|---|---|---|
-| 1. 2048 from-scratch | 125k | 0.2755 | 0.2812 |
-| 2. 4096 finetune | 135k | 0.3557 | 0.3510 |
-| 3. Endpoint cooldown | 170k | 0.3716 | 0.3665 |
-
-The stage 1 -> stage 2 jump (+0.08 HSS on 4096) is the biggest single improvement and motivates the 2048 -> 4096 transfer. Stage 3 (endpoint cooldown) adds another +0.016. Note how stage 1 is slightly better at 2048 than at 4096 (because it was only trained on 2048), while stages 2 and 3 invert that ordering after being finetuned on 4096.
-
-### Previously submitted model (2048, single-stage)
-
-The `submitted_2048/` directory holds the checkpoint we actually sent to the public leaderboard. It was trained in a single stage on 2048-point data and is a direct ancestor of the current release.
-
-| Split | Metric | Score |
-|---|---|---|
-| **Public leaderboard (test)** | **HSS** | **0.427** |
-| Internal val @ 2048 | HSS | 0.3692 |
-| Internal val @ 4096 | HSS | 0.3665 |
-
-We do not have a test number for the current release, but the val-to-test gap observed on this 2048 submission was about **+0.06 HSS** (0.37 val -> 0.43 test). A similar gap on the current `checkpoint.pt` (0.382 val) would suggest a test score in the low 0.44s, though this is extrapolation and unverified.
-
-## Reproducibility
-
-| Test | Result |
-|---|---|
-| Forward pass (same ckpt, same input) | bit-identical (0.00 diff) |
-| Deterministic mode, 3 independent runs | bit-identical (162 tensors, max_diff=0.0) |
-| Step 3 from same stage-2 ckpt (2 runs) | HSS=0.382, 0.384 |
-| Compiled-mode E2E variance across runs | ~0.03 HSS (Triton kernel nondeterminism) |
-
-`reproduce_deterministic.sh` produces byte-identical weights across runs with the same seed, at the cost of ~2x slower training (no `torch.compile`). Compiled mode has small run-to-run variance from Triton kernel selection that grows through chaotic SGD dynamics; E2E compiled repros land in the 0.349-0.379 range.
-
-A subtle iteration-order effect: the shipped `bad_samples.txt` has 156 non-empty entries (the file lacks a trailing newline so `wc -l` reports 155). Two additional bad samples were discovered after training - they are legitimately bad GT but adding them changes the batch iteration order and costs ~0.005 HSS in deterministic mode and ~0.04 in compiled mode. See the "Reproducibility Notes" section of `REPRODUCE.md` for the full story.
-
-## Repository layout
+## Layout
 
 ```
-checkpoint.pt                    shipped HSS=0.382 model (step 170000), 4096-point input
-script.py                        competition inference entry point (uses checkpoint.pt)
-s23dr_2026_example/              training package (model, data, train loop, losses)
-configs/base.json                shared training config
-reproduce.sh                     compiled-mode E2E reproduction (~3 hr)
-reproduce_deterministic.sh       bit-reproducible E2E reproduction (~5.5 hr)
-REPRODUCE.md                     detailed recipe, results, ablations, notes
-
-submitted_2048/                  the model we actually sent to the public leaderboard (HSS_test=0.427)
-    checkpoint.pt                single-stage 2048 model (step 160000)
-    args.json                    full training args
-    README.md                    training details and val/test scores
-
-repro_runs/                      evidence that the 3-stage recipe reproduces
-    e2e_repro4_hss379/           closest compiled E2E repro (val HSS=0.374)
-    compiled_repro_hss376/       best compiled repro from this codebase (val HSS=0.376)
-    deterministic_hss372/        bit-reproducible deterministic repro (val HSS=0.372)
+script.py                    Submission entry: load checkpoint, run pipeline,
+                             write submission.json.
+checkpoint.pt                Trained weights (~106 MB, Git LFS). Dev val
+                             HSS=0.382 (original training run; not exactly
+                             reproducible). Public test HSS=0.4470.
+configs/base.json            Shared training config (architecture + optimizer).
+reproduce.sh                 3-stage retraining (torch.compile, fast).
+reproduce_deterministic.sh   Bit-reproducible variant (no compile, ~2x slower).
+make_datasets.sh             Rebuild sampled datasets from raw hoho22k tars.
+REPRODUCE.md                 Recipe, ablations, reproduction numbers.
+
+s23dr_2026_example/          The pipeline as an importable package.
+  point_fusion.py            Per-view fusion -> labeled 3D point cloud.
+  cache_scenes.py            Stream raw shards -> per-scene .pt files.
+  make_sampled_cache.py      .pt -> fixed-size .npz priority samples.
+  data.py                    Dataset wrapper + augmentation.
+  tokenizer.py               Per-point tokens (xyz + Fourier + class + src
+                             + behind + vote features).
+  model.py                   Perceiver encoder + segment decoder.
+  attention.py               SDPA blocks with optional QK-norm.
+  losses.py                  Sinkhorn matching + endpoint L1 + confidence BCE.
+  sinkhorn.py                Optimal-transport segment matcher.
+  varifold.py                Segments -> vertices/edges; varifold loss kernels.
+  wire_varifold_kernels.py   Varifold kernels (alternate loss, unused at 0.382).
+  segment_postprocess.py     Iterative vertex merging.
+  postprocess_v2.py          Snap-to-cloud and horizontal-edge alignment.
+  color_mappings.py          ADE20K + Gestalt label palettes.
+  train.py                   Training loop driven by reproduce.sh.
+  bad_samples.txt            156 scenes excluded from training (misaligned GT).
+
+repro_runs/                  Three frozen reproduction runs:
+  compiled_repro_hss376/       best compiled run from this codebase
+  e2e_repro4_hss379/           closest end-to-end repro to original
+  deterministic_hss372/        bit-reproducible run
+
+submitted_2048/              The earlier public-leaderboard checkpoint
+                             (2048-only, public test HSS=0.4273). The current
+                             top-level checkpoint.pt scored 0.4470 on test.
 ```
 
-Each directory under `repro_runs/` contains the three stage-final checkpoints (125k / 135k / 170k) plus their `args.json`, so a participant can resume from any stage. Note the directory names carry the score at the time the directory was created, which may differ by ~0.002 from fresh evals in the table above due to random variation in post-processing and CUDA kernel selection.
-
-## Related branches
-
-- `main` - this release
-- `best-4096-transfer` - working branch with full commit history and internal dev notes
-- `validation-archive` - cold archive of all validation runs (logs, final checkpoints, args) used to verify the release
-
-## License
-
-**CC-BY-NC 4.0.** The model weights and code in this repository are released under the Creative Commons Attribution-NonCommercial 4.0 International license. You are free to use, share, and adapt this work for **non-commercial** purposes, provided you give appropriate **attribution**. The training and validation datasets (`usm3d/s23dr-2026-sampled_*`) have their own terms - see the S23DR 2026 competition page for details.
-
-## Acknowledgements
-
-This checkpoint is released as a public learned baseline for participants of the **S23DR 2026** challenge, part of the [USM3D workshop](https://usm3d.github.io) at CVPR 2026.
+## Submissions and scores
+
+Two submissions have been evaluated on the public test set. Both are kept in
+this repo. "Dev val" below is the last 1024 scenes of the published training
+set, which is what we optimized against; we did not eval on the official
+validation split. See "Evaluation sets" in REPRODUCE.md for the formal
+definitions.
+
+| Model | Path | script.py settings | Dev val HSS | **Public test HSS** |
+|---|---|---|---:|---:|
+| 2048 (single-resolution, earlier) | `submitted_2048/checkpoint.pt` | `SEQ_LEN=2048`, `CONF_THRESH=0.7`, single-pass merge | 0.369 @ 2048 | **0.4273** |
+| 4096 (3-stage transfer, current)  | `checkpoint.pt`                 | `SEQ_LEN=4096`, `CONF_THRESH=0.5`, iterative merge   | 0.382 @ 4096 | **0.4470** |
+
+The 4096 model is the better submission on both dev val (+0.013) and public
+test (+0.020). The leaderboard is open; the test numbers above are what each
+checkpoint scored when submitted at its corresponding commit (`f4487da` for
+2048, `4946666` for 4096). Note that *three* things change between the two
+submissions - the model, the inference `SEQ_LEN`, and the merge / confidence
+postprocessing - so the +0.020 test gain is not attributable to the model
+alone.
+
+## Notes
+
+- The model predicts line *segments* in per-scene-normalized coordinates
+  (parametrized as midpoint + half-vector, decoded as endpoint pairs). The
+  script multiplies by the scene scale and adds the center to lift them to
+  world space before extracting vertex/edge lists; merging and snapping
+  postprocessing then run in world space. Output vertices are in the same
+  world frame as the COLMAP / ground-truth reconstruction.
+- The pipeline returns a 2-vertex / 1-edge dummy wireframe for any sample
+  where fusion fails or no segment passes the confidence threshold.
+- There is (unfortunately) a fair amount of run-to-run variation: both
+  *interrun* (same seed, different runs of `reproduce.sh` give dev val HSS
+  in 0.342-0.379 due to `torch.compile` picking different Triton kernels -
+  see REPRODUCE.md) and *interseed* (varying `--seed` gives non-trivially
+  different final dev val HSS as well).
+
+
+## Ideas / Suggestions / Thoughts
+
+- There is a messy repo with many experiments and a codebase that has more features, but is messier [here](https://github.com/JackLangerman/s23dr_2026_example)
+- The training protocol, initialization, and architecture could probably all be tweaked to improve stability / interseed variation. Increasing sinkhorn eps at the beginning of training, and longer warmup both seemed to yield less variation between seeds, but worse absolute performance; can we improve both?
+- Can training on Building World help? I tried a single "pretraining" run on BuildingWorld and it didn't help, but maybe training on both simultaneously would be better? 
+- What about training on image (Gestalt/ADE) patches + rays instead of fused points? Can you do better with less inductive bias?
+- What about mixed 2048 / 4096 training (instead of sequential)? does it help?
+- What about other loss functions? a better softHSS? 
+- Does a bigger model with more data (eg from building world) help? what about with better / different regularization or data augmentation?
+- What about other modern DETR tricks?
+- Can you output a graph directly instead of post-hoc vertex merging?
+- bad_samples.txt doesn't include all the bad samples, but in quick tests, adding the two bad samples shown in REPRODUCE.md made scores worse. Can you make the score better while excluding these?
+
+### I'm sure there is a ton more cool stuff to do! Be creative! Excited to see what you come up with!

REPRODUCE.md

@@ -28,6 +28,32 @@ Perceiver: hidden=256, ff=1024, latent_tokens=256, latent_layers=7
 
 All shared config lives in `configs/base.json`.
 
+## Evaluation sets
+
+Three distinct evaluation sets show up in this work. Every HSS / F1 / IoU
+number below is from one of these three; we try to tag each number with which.
+
+- **Dev val** = the last 1024 scenes of the published training set
+  (`hf://usm3d/s23dr-2026-sampled_*_v2:train`). This is what we actually
+  optimized against during development, and it is the set behind every
+  "HSS=0.382"-style number in this document, in `submitted_2048/README.md`,
+  and in the run-history files under `repro_runs/` and the validation-archive.
+- **Official validation** = `hf://usm3d/s23dr-2026-sampled_*_v2:validation`
+  (equivalently the `*public*` tars in `usm3d/hoho22k_2026_test_x_anon`).
+  We did *not* eval on this split during development. No HSS number in this
+  repo refers to it.
+- **Public test** = the `*private*` tars in `usm3d/hoho22k_2026_test_x_anon`,
+  scored by the competition harness and posted to the leaderboard. We have
+  two such numbers, both clearly labeled "public test" wherever they appear:
+  **0.4273** (2048 submission, commit `f4487da`) and **0.4470** (4096
+  submission, commit `4946666`).
+
+Because we never validated against the official validation split, there is
+some risk that the dev-val numbers are mildly overfit to the last-1024-train
+slice. The +0.06 dev-val-to-public-test gap (consistent across both
+submissions, see `submitted_2048/README.md`) is empirically positive, but it
+is not a substitute for actually scoring on official val.
+
 ### Step 1: 2048 Phase 1 (from scratch) — ~1.5hr
 
 ```
@@ -49,7 +75,7 @@ optimal transport loss learns to match predicted segments to ground truth.
 **Why 2048 first:** Training directly on 4096 overfits (1.47x train/val ratio
 vs 1.19x for 2048). The 2048 model learns better-generalized representations.
 
-**Output:** HSS ~0.28.
+**Output:** dev val HSS ~0.28.
 
 ### Step 2: 4096 finetune (constant LR) — ~15min
 
@@ -66,9 +92,9 @@ Switches input from 2048 to 4096 points, increasing structural coverage from
 66% to 74%. The gentle lr (3e-5) preserves learned representations while
 adapting to the extra input. Higher LR (>1e-4) causes catastrophic forgetting.
 
-HSS jumps from 0.28 to 0.35 in ~5k steps. Plateaus by 10k steps.
+Dev val HSS jumps from 0.28 to 0.35 in ~5k steps. Plateaus by 10k steps.
 
-**Output:** HSS ~0.35.
+**Output:** dev val HSS ~0.35.
 
 ### Step 3: Cooldown with endpoint loss — ~1hr
 
@@ -86,12 +112,18 @@ Adds symmetric endpoint L1 loss (using detached sinkhorn assignment) to
 tighten vertex precision. The sinkhorn loss alone operates on segment
 midpoint/direction/length and doesn't directly penalize endpoint position error.
 
-**Output:** HSS=0.382, F1=0.414.
+**Output:** dev val HSS=0.382, F1=0.414. Public test HSS=0.4470.
 
 ### Key Numbers
 
-| Stage | Steps | HSS | F1 | What changed |
-|-------|-------|-----|-----|-------------|
+Per-stage **dev val** scores from the **original training run** (March 23-26),
+which produced the shipped `checkpoint.pt`. Re-running `reproduce.sh` from
+scratch does not hit these exact numbers - see "Reproduction Results" below for
+the actual ranges. Compiled-mode re-runs of Step 3 land in dev val 0.342-0.379,
+with the best run from this codebase at 0.376.
+
+| Stage | Steps | Dev val HSS | Dev val F1 | What changed |
+|-------|-------|-------------|------------|--------------|
 | After Step 1 | 125k | 0.281 | 0.156 | Learned geometry from 2048 pts |
 | After Step 2 | 135k | 0.351 | 0.190 | +74% coverage from 4096 pts |
 | After Step 3 | 170k | **0.382** | **0.411** | Vertex precision from endpoint loss |
@@ -111,53 +143,78 @@ midpoint/direction/length and doesn't directly penalize endpoint position error.
    directly penalizes vertex position errors, which sinkhorn loss alone
    doesn't do (it operates on midpoint/direction/length parametrization).
 
-## What Doesn't Work
-
-- **Training 4096 from scratch:** overfits (1.47x train/val gap), peaks at 0.346
-- **BuildingWorld pretraining:** representations are orthogonal to S23DR (cosine sim = 0.05)
-- **Mixed BW+S23DR training:** BW data hurts due to domain gap
-- **High dropout / weight decay:** prevents overfitting but causes underfitting
-- **High finetune LR (>1e-4):** catastrophic forgetting of 2048 representations
-- **Steeper cooldown (1e-5, 20x drop):** slightly worse than 3e-5 for this checkpoint
+## What Doesn't Work (yet)
+
+These are informal observations from one-off experiments during development.
+The runs, args, and eval logs are mostly [here](https://github.com/JackLangerman/s23dr_2026_example),
+but not all of them are preserved perfectly. The specific numbers below come from contemaranious notes, 
+but are not all trivially reproducable. Take them as directional guidance, not as benchmarks.
+
+- **Training 4096 from scratch:** observed to overfit (~1.47x train/val loss
+  gap, vs ~1.19x for 2048) and peak around dev val HSS 0.346 in a single run.
+- **BuildingWorld pretraining:** in one experiment, the representations were
+  near-orthogonal to S23DR (cosine sim ~0.05 between learned features) and
+  did not transfer.
+- **Mixed BW+S23DR training:** mixing BW data into the S23DR loader hurt
+  dev val HSS in the runs we tried, presumed to be from domain gap.
+- **High dropout / weight decay:** lowered the train/val gap but also lowered
+  dev val HSS in the configurations we tried.
+- **High finetune LR (>1e-4):** dropped dev val HSS sharply in Step 2 in
+  single-run observations, consistent with disrupting the Step 1 representations.
+- **Steeper cooldown (1e-5, 20x drop):** slightly worse than 3e-5 in the one
+  comparison we ran for this checkpoint.
 
 ## Reproduction Results
 
 ### End-to-end reproductions
 
-| Model | HSS | F1 | IoU | Notes |
-|-------|-----|-----|-----|-------|
-| Original | 0.382 | 0.414 | 0.370 | Shipped checkpoint |
+All HSS / F1 / IoU below are on **dev val**.
+
+| Model | Dev val HSS | Dev val F1 | Dev val IoU | Notes |
+|-------|-------------|------------|-------------|-------|
+| Original | 0.382 | 0.414 | 0.370 | Shipped checkpoint, original training run, not reproducible from this codebase |
 | E2E repro #4 | 0.379 | 0.409 | 0.369 | Closest E2E, `repro_runs/e2e_repro4_hss379/` |
 | Compiled repro (from submission codebase) | 0.376 | — | — | Best compiled repro from this codebase, `repro_runs/compiled_repro_hss376/` |
 | E2E repro #3 | 0.375 | 0.404 | 0.367 | |
 | Deterministic E2E | 0.372 | 0.398 | 0.368 | Bit-reproducible, `repro_runs/deterministic_hss372/` |
-| E2E repro #5 | 0.349 | 0.373 | — | Outlier (early compile divergence) |
+| E2E repro #5 | 0.349 | 0.373 | — | Compiled, low end of cluster |
+| `reproduce.sh` smoketest | 0.342 | — | — | Single run of the published script end-to-end (validation-archive `runs/reproduce_smoketest/`) |
 
 ### Partial reproductions (isolating pipeline stages)
 
-| Test | Starting from | HSS | Gap to original |
-|------|--------------|-----|-----------------|
+| Test | Starting from | Dev val HSS | Gap to original |
+|------|--------------|-------------|-----------------|
 | Step 3 from orig Step 2 (run A) | Original step135000.pt | 0.382 | 0.000 |
 | Step 3 from orig Step 2 (run B) | Original step135000.pt | 0.384 | +0.002 |
 | Step 2+3 from orig Step 1 | Original step125000.pt | 0.377 | -0.005 |
-| Step 1 from orig step 100k | Original step100000.pt | 0.285 (Step 1 HSS) | +0.004 vs 0.281 |
+| Step 1 from orig step 100k | Original step100000.pt | 0.285 (Step 1) | +0.004 vs 0.281 |
 
-Step 3 from the same checkpoint reproduces to within 0.002. The E2E variance
-(0.349-0.379) is dominated by torch.compile nondeterminism in Step 1.
+Step 3 from the same checkpoint reproduces to within 0.002 dev val HSS. The
+full E2E dev val HSS variance (0.342-0.379, see All benchmarks below) is
+dominated by torch.compile nondeterminism in Step 1.
 
 ### All benchmarks
 
-| Model | Input | HSS | F1 | IoU | Notes |
-|-------|-------|-----|-----|-----|-------|
+The HSS / F1 / IoU columns below are all on **dev val**. Public-test scores
+appear in the Notes column where available.
+
+| Model | Input | Dev val HSS | Dev val F1 | Dev val IoU | Notes |
+|-------|-------|-------------|------------|-------------|-------|
 | Handcrafted baseline | raw views | 0.307 | 0.404 | 0.260 | |
-| h256+qk+ep (submitted) | 2048 | 0.365 | 0.388 | 0.360 | HSS=0.427 on test |
+| h256+qk+ep (submitted) | 2048 | 0.365 | 0.388 | 0.360 | Public test HSS=0.4273 (commit f4487da) |
 | Original 3-step | 2048 | 0.373 | 0.404 | 0.363 | |
-| Original 3-step | 4096 | 0.382 | 0.414 | 0.370 | Best ever |
-| Step3 repro from orig S2 | 4096 | 0.384 | 0.414 | — | Near-exact repro |
+| Original 3-step | 4096 | 0.382 | 0.414 | 0.370 | Best ever, original training. **Public test HSS=0.4470** (commit 4946666) |
+| Step3 repro from orig S2 | 4096 | 0.384 | 0.414 | — | Near-exact repro from a saved Step 2 ckpt |
 | E2E repro #4 | 4096 | 0.379 | 0.409 | 0.369 | |
 | Compiled repro (submission codebase) | 4096 | 0.376 | — | — | Best compiled from this exact codebase |
 | E2E repro #3 | 4096 | 0.375 | 0.404 | 0.367 | |
-| Deterministic E2E | 4096 | 0.372 | 0.398 | 0.368 | Bit-reproducible |
+| Deterministic E2E | 4096 | 0.372 | 0.398 | 0.368 | Bit-reproducible across runs (different trajectory than compiled) |
+| E2E repro #5 | 4096 | 0.349 | 0.373 | — | Compiled, low end of cluster |
+| `reproduce.sh` smoketest | 4096 | 0.342 | — | — | Single E2E run of the published `reproduce.sh` |
+
+The shipped 0.382 is by definition not reachable from this codebase (the
+original training run is gone). Best compiled repro is **0.376**, mode of
+compiled repros is around 0.375-0.379, and the lower tail extends to 0.342.
 
 ## Code Equivalence Verification
 
@@ -167,28 +224,40 @@ Step 3 from the same checkpoint reproduces to within 0.002. The E2E variance
 | Loss computation | Bit-identical (0.00 diff) |
 | Gradient computation | 5e-8 max diff |
 | Training from same seed | Bit-identical steps 1-44 |
-| Step 3 from same checkpoint (2 runs) | HSS=0.382, 0.384 |
+| Step 3 from same checkpoint (2 runs) | Dev val HSS=0.382, 0.384 |
 | Deterministic mode (2 runs) | Bit-identical (0.00 diff) |
 
 ## Reproducibility Notes
 
+All HSS numbers in this section are on **dev val**.
+
 **Default mode** (`reproduce.sh`): Uses torch.compile (~3x faster). Each run
 gets different Triton kernels, causing ~1e-8 floating-point divergence at a
-random step (31-45). This grows through chaotic SGD dynamics, giving HSS
-variance of ~0.03 across runs. E2E reproductions land in the 0.349-0.379 range.
-
-**Deterministic mode** (`--deterministic` flag): Disables torch.compile.
-Bit-identical across runs with the same seed. HSS=0.372 (slightly lower than
-compiled mode because eager-mode kernels follow a different numerical path).
+random step (31-45). This grows through chaotic SGD dynamics. Documented
+compiled E2E runs from this codebase have produced dev val HSS in
+**0.342-0.379**; the modal cluster is 0.375-0.379, with low-side excursions
+at 0.349 and 0.342 (no high-side outliers, so the distribution is one-sided
+rather than symmetric around a mean). Best compiled repro is 0.376, vs the
+shipped 0.382 from the original (lost) training run.
+
+**Deterministic mode** (`--deterministic` flag): Disables torch.compile and
+forces CUDA deterministic ops. Bit-identical across runs with the same seed
+(verified across 3 independent runs). Dev val HSS=0.372. Note: deterministic
+mode **diverges from compiled mode at step 1** because eager and compiled
+forward passes use different floating-point reduction orders - it is a
+different numerical trajectory entirely, not a reproduction of any compiled run.
 
 **bad_samples.txt**: The shipped file has 156 entries to match original training.
 (Note: `wc -l` reports 155 because the last line lacks a trailing newline.)
 Two additional bad samples (`47b0e0ce19b`, `4b2d56eb3ef`) were discovered after
 the original training run. They are legitimately bad (misaligned GT) but were
 included in the original training data. Adding them changes the batch iteration
-order and costs ~0.005 HSS in deterministic mode (0.372 -> 0.367) and ~0.04 in
-compiled mode due to compounded torch.compile variance. Participants training
-from scratch may wish to add these 2 entries for cleaner training data, but
-should expect slightly different scores due to the changed iteration order.
-
-The shipped `checkpoint.pt` is from the original training run (HSS=0.382).
+order and costs ~0.005 dev val HSS in deterministic mode (0.372 -> 0.367) and
+~0.04 in compiled mode (0.376 -> 0.335 in our `validate_155_compiled` vs
+`validate_158_compiled` runs) due to compounded torch.compile variance.
+Participants training from scratch may wish to add these 2 entries for cleaner
+training data, but should expect slightly different scores due to the changed
+iteration order.
+
+The shipped `checkpoint.pt` is from the original training run
+(dev val HSS=0.382, public test HSS=0.4470).

submitted_2048/README.md

@@ -1,16 +1,19 @@
-# Submitted 2048 Model (public leaderboard entry)
+# Submitted 2048 Model (earlier public leaderboard entry)
 
-This is the checkpoint that was actually submitted to the S23DR 2026 public leaderboard. It trains on the 2048-point dataset only (single-stage, no 4096 transfer). The current top-level `checkpoint.pt` (HSS=0.382 val) is its direct descendant via the 3-step 2048 -> 4096 -> endpoint-cooldown recipe.
+This is the **earlier** of two checkpoints submitted to the S23DR 2026 public leaderboard. It trains on the 2048-point dataset only (no 4096 transfer); within 2048 it is two-phase (phase 1 from scratch + phase 2 with endpoint loss and cooldown, resumed from `step125000.pt`). The current top-level `checkpoint.pt` (dev val HSS=0.382, public test HSS=0.4470) is its descendant via the 3-step 2048 -> 4096 -> endpoint-cooldown recipe and is the better submission on both dev val and public test.
 
 | Split | Metric | Score |
 |---|---|---|
-| Public leaderboard (test) | HSS | **0.427** |
-| Internal val (2048, 1024 samples) | HSS_conf | 0.369 |
-| Internal val (4096, 1024 samples) | HSS_conf | 0.367 |
+| Public test (leaderboard) | HSS | **0.4273** |
+| Dev val (last 1024 train scenes), 2048-pt input | HSS_conf | 0.369 |
+| Dev val (last 1024 train scenes), 4096-pt input | HSS_conf | 0.367 |
+
+We did not eval on the official validation split (`hf://usm3d/s23dr-2026-sampled_*_v2:validation`)
+during development, so no number here refers to it. See "Evaluation sets" in `../REPRODUCE.md`.
 
 ## Training details
 
-Single-stage training on `hf://usm3d/s23dr-2026-sampled_2048_v2:train`:
+Two-phase 2048-only training on `hf://usm3d/s23dr-2026-sampled_2048_v2:train` (phase 1 from scratch to step 125k, phase 2 resumed from `step125000.pt` and trained to step 160k with endpoint loss and a 20k-step cooldown ending at step 160k):
 
 - **Architecture:** same Perceiver as the current release (hidden=256, latent_tokens=256, latent_layers=7, segments=64)
 - **Input:** 2048 points
@@ -30,6 +33,13 @@ This checkpoint expects 2048-point input. To run it with the submission harness
 
 ## Why it is included
 
-The current release (`../checkpoint.pt`, HSS=0.382 val) is a strict improvement over this one, but only on the internal val split. The **0.427 public leaderboard score** is the only test-set number we have, so this checkpoint is preserved as the empirical anchor for the val-to-test gap.
+The current release (`../checkpoint.pt`, dev val HSS=0.382) is the better submission on both dev val and public test. This older checkpoint is preserved as the empirical anchor for the dev-val-to-public-test gap.
+
+Dev-val-to-public-test gap observed across both submissions:
+
+| Submission | Dev val HSS | Public test HSS | Gap |
+|---|---|---|---|
+| 2048 (this checkpoint) | 0.369 | 0.4273 | +0.058 |
+| 4096 (`../checkpoint.pt`) | 0.382 | 0.4470 | +0.065 |
 
-Val-to-test gap observed: **0.369 val -> 0.427 test** (about +0.06). The same train/val/test relationship should roughly carry over to the current 0.382-val release, but we do not have a test number for it since the leaderboard uses this older model.
+Both submissions show roughly the same +0.06 dev-val-to-public-test gap, so dev val HSS appears to be a reasonable proxy for public test HSS at this scale (subject to whatever distributional differences exist between the dev val split, the official validation split we did not eval on, and the public test split). Note that the inference pipeline also changed between the two submissions (`SEQ_LEN` 2048 -> 4096, `CONF_THRESH` 0.7 -> 0.5, single-pass merge -> iterative merge), so the +0.020 public test gain is not attributable to the model alone.