SparDA checkpoint artifacts
Learned-selector checkpoints (indexer / route-projector weights) from the SparDA research runs. These are not MiniCPM backbone weights and cannot run as a standalone model.
On 2026-10-13 this repository was split: the published weights stayed here, and the research
workspace records (.hl/, .planning/), result files and analyses moved to
Xnhyacinth/sparda-research-private
(private). The repository was previously named sparda-artifacts; Hugging Face redirects the
old name here.
| Files | 99 .pt |
| Bytes | 2,209,451,107 (2209.5 MB) |
| Loading | torch.load(path, map_location='cpu', weights_only=True) |
| Inventory | MANIFEST_weights.json |
Companion collection: Attention-Checkpoints.
Why the weights are re-exported
Every file here was re-saved to strip absolute workspace paths and account names embedded in
its stored training arguments. Each tensor was compared with torch.equal after reload
(27,362 tensors across 99 files), so the tensor payload is
identical to the original — only string values differ. Because the file bytes changed, the
checkpoint_sha256 values recorded in older result files do not match the published bytes;
PROVENANCE_weights_redaction.json maps each original
SHA-256 to its published SHA-256.
Contents
| Directory | Files | MB | What it is |
|---|---|---|---|
paper1/ |
12 | 50.9 | Selected pure deployment exports from fixed candidates and five ablation runs |
car_consultation_20260928/ |
18 | 76.5 | CAR consultation ablation pilots, 18 checkpoints |
nosa_paper1_20260928/ |
3 | 409.2 | NoSA paper-1 run, 3 checkpoints |
qams_20260915/ |
10 | 400.9 | QAMS candidate sweep 2026-09-15 (2 backbone-scale runs, 6 block-scale candidates, 2 tiny) |
qams_20260919/ |
4 | 8.9 | QAMS joint/dual-pageid pilot 2026-09-19 |
qams_car_formal/ |
2 | 0.8 | CAR formal 4-GPU run (best-accuracy + step 100) |
qams_car_v11/ |
2 | 5.2 | CAR v1.1 diagnostic (best-accuracy + step 100) |
qams_d/ |
1 | 65.0 | QAMS-D forecast bundle |
qams_fs/ |
2 | 390.4 | QAMS-FS formal run (latest complete step + best in-loop accuracy) |
sync_snapshot_20260930/ |
45 | 801.6 | Snapshot of the training workspace checkpoint directories, 45 files (includes the sole copy of the qams_formal step-880 and nosa_sparda ckpt-1950/1800 payloads) |
The seven deployment-scale indexers are stored as individually named files:
qams_fs/ckpt-600.pt— 195,199,873 bytesqams_fs/ckpt-best-acc74.4381-step000540.pt— 195,204,903 bytesqams_car_v11/ckpt-100.pt— 2,578,481 bytesqams_car_v11/ckpt-best-acc26.4607-step000060.pt— 2,587,655 bytesqams_car_formal/ckpt-100.pt— 396,081 bytesqams_car_formal/ckpt-best-acc27.8854-step000060.pt— 400,007 bytesqams_d/qams-pd-forecast680-r1-softmax.pt— 65,041,177 bytes
Every other weight is a training-run checkpoint under the directories above; the full list of 99 paths with sizes and SHA-256 is in MANIFEST_weights.json.
Selected Paper1 deployment exports
paper1/ contains the fixed MiniCPM step-140 candidate, NOSA final step-300 candidate, and final/original-selected-best exports for five MiniCPM ablation runs. These are selector deployment weights; selection does not establish benchmark superiority. Optimizer state, intermediate checkpoints, results, and research documents are excluded. Original-to-published checksums and tensor equality checks are recorded in the provenance manifest. The curated Attention-Checkpoints subset remains separate; private research evidence is stored in the existing private research repository.