Nanberry β private checkpoint backup
Paired curriculum pilot completed; neither arm promoted. Verified negative results. Both512-update arms finished in36m39s total. No supported transfer advantage; retention regressions. These are experimental backups, not a default change.
Completed mixed-v2 continuation: one full55,043-row/11,235-update pass from mixed-v1. Reused matched macro83.94% versus cached Jev70.19%; grounding59β117/148. Money regresses43β34/77; fresh-world scheduling remains weak. No default promotion, extra epoch, paid teacher requests, output postprocessor or Confirmation evaluation.
Seventeen snapshots: prior thirteen preserved, plus256/512 snapshots of two independent paired-pilot arms from mixed-v2. The mixed-v1 full epoch is complete; all three128-update pilot candidates failed their preset gates. LoRA + decision head, not a standalone chat model.
Tiny-fit LR diagnostic: two independent128-update arms, same51 repeated training examples and51 different-world controls. Higher LR fits51/51 versus29/51 at the repair LR, but controls remain26/51 versus27/51. Tiny-set fitting is possible; reliable transfer is not demonstrated. Neither arm is promoted or authorized for continuation. Output surgery remains rejected.
Task-specific pilots: independent grounding+replay and scheduling+replay,128updates each from the unchanged mixed-v1 parent. Grounding deeper supported recall stays3/24; scheduling-focused balanced accuracy55β50.625%, false accepts25β43/80. Both rejected; no extension/default promotion. Reserved Confirmation remains unmeasured.
Bounded repair pilot: exactly128updates/498rows/203,647tokens. Grounding overconfidence improved, but deep supported recall only3β4/24, scheduling balanced accuracy remained55%, and manual/procedural-soft retention gates failed. Rejected; no extension or default promotion. Confirmation remains unmeasured.
Mixed-epoch results: sampled hard-target Test accuracy 72.5% β 88.8%, OOD 48.6% β 65.4%, versus the Open-Jev parent. Grounding became much more overconfident; OOD scheduling and small historical retention accuracy declined. Not promoted as the default.
Prior Open-Jev results: sampled Test hard-target accuracy 75.9% β 95.0%, OOD 77.7% β 93.4%, versus the immediate parent. Procedural-retention loss worsened 17.7%; this is not an unconditional all-round promotion or a fresh comparison with the untouched 5090 checkpoint.
| Snapshot | Exactly what it contains |
|---|---|
| Mixed-v2 FULL EPOCH | Independent mixed-v1 continuation:11,235updates/55,043rows/17,181,010tokens. Full resume state. Stronger overall reused-panel result, documented money/scheduling limits; NOT promoted. |
| 5090 FULL EPOCH | Original, untouched 5090 result: one full SargeDev-v3 epoch, 655,806 records, 82,835 updates. No later procedural or Open-Jev-v1.1 training. Includes optimizer/RNG/resume state. |
| General parent | The 5090 result plus 480 procedural-training updates. This is the starting point for Open-Jevβnot the untouched full-epoch result. LoRA/head weights only; no optimizer. |
| Open-Jev IN PROGRESS | General parent plus 815 / 34,795 Open-Jev updates (2.34%). Includes resume state. Not a completed epoch or validated improvement. |
| Open-Jev FULL EPOCH | General parent plus all 34,795 Open-Jev updates / 147,139 Train rows. Includes resume state. Completed final evaluations; see the gains and retention warning above. |
| Mixed-v1 FULL EPOCH | Open-Jev full epoch plus 10,583 updates / 55,039 mixed Train rows, 16,022,005 useful tokens. Fresh optimizer for this completed epoch; final resume state retained. Gains and material grounding-confidence regression documented above. |
| Mixed repair PILOT | Mixed-v1 parent plus exactly128 updates on498 of2048 candidate Train rows. Full optimizer/RNG state; not a full epoch, not promoted, and rejected by four preset validation/retention gates. |
| Grounding task PILOT | Independent restart from mixed-v1;128updates/548rows/196,928tokens, grounding+replay. Four preset gates failed; not promoted. |
| Scheduling task PILOT | Independent restart from mixed-v1;128updates/566rows/198,618tokens, scheduling+replay. Four preset gates failed; not promoted. |
| Tiny-fit repair_lr, step64 | Midpoint of an independent128-update optimization diagnostic;51 unique Train examples repeated, no replay. Optimizer/RNG retained; no default promotion. |
| Tiny-fit repair_lr, step128 | Final of an independent128-update optimization diagnostic;51 unique Train examples repeated, no replay. Optimizer/RNG retained; no default promotion. |
| Tiny-fit full_epoch_lr, step64 | Midpoint of an independent128-update optimization diagnostic;51 unique Train examples repeated, no replay. Optimizer/RNG retained; no default promotion. |
| Tiny-fit full_epoch_lr, step128 | Final of an independent128-update optimization diagnostic;51 unique Train examples repeated, no replay. Optimizer/RNG retained; no default promotion. |
Shared ancestry: 5090 full epoch β procedural +480 β Open-Jev +34,795 β mixed-v1 +10,583. From mixed-v1, five separate128-update experiments branch: mixed repair, grounding+replay, scheduling+replay, and the two tiny-fit LR diagnostics. Each tiny-fit arm has a64-step midpoint and128-step final snapshot. They are siblings, not cumulative continuations; the +815 Open-Jev snapshot remains an earlier intermediate backup. No maze-specialized checkpoints are included. Snapshots are fixed, not automatically synchronized.
Datasets and licenses
Read DATASETS-AND-LICENSES.md. It identifies each dataset, exact revision, training scope, label provenance, upstream licenses and unresolved restrictions. Original publisher cards/notices are preserved under licenses/. There is no blanket MIT/Apache license or commercial-clearance claim for this mixed-provenance collection.
Code and formats
- scripts/train_jev.py: training and strict checkpoint restoration.
- src/nb_shim.py + third_party/: backbone adapter, decision head and required vendored modules/license.
- scripts/openjev.py: pinned intake/configuration and CPU checks.
- configs/: exact configs; manifest.json: sizes, SHA-256 and lineage.
All checkpoints contain a trainable state dictionary; obtain the pinned Nanbeige backbone separately. The original 5090 epoch used mixed NVFP4/FP8/BF16; subsequent stages use BF16, fixed depth 2. BF16 loading does not reproduce the original mixed arithmetic. Generic train_jev.py deliberately refuses unsupported mixed-precision restoration; these are archival artifacts, not a turnkey 5090 environment.
Resume the historical incomplete Open-Jev snapshot
The completed Open-Jev checkpoint is the Open-Jev FULL-EPOCH file above; the authorized one-epoch runs are already finished. The following command is only for reproducing continuation from the older step-815 backup, in a separate clean workspaceβnot for extending or overwriting the completed run.
Requires the original data workspace, split audits and retention panels; datasets are not bundled. Never overwrite a newer local checkpoint.
pip install -r requirements.txt
mkdir -p runs/nanberry-openjev-v11-epoch1
cp checkpoints/openjev-IN-PROGRESS-step-00000815-resume.pt runs/nanberry-openjev-v11-epoch1/latest.pt
python scripts/train_jev.py --config configs/openjev-v11-full-epoch.json --resume
No frozen base weights, raw input datasets, credentials or caches are included. results/ contains evaluation IDs, label/probability vectors and audit summaries; original input text is excluded.
Mixed-v2 is a sixth independent branch from mixed-v1, not a continuation of the rejected pilots or overfit tiny-set checkpoints. Fresh and historical baseline probabilities differ slightly because BF16 pooling is not bitwise reproducible; the explicit replay amendment and both baselines are preserved.
Model tree for raspbfox/nanberry-decision-checkpoints
Base model
Nanbeige/Nanbeige4.2-3B-Base