Register exact Graph Student HARD full continuation launch
Browse files
student_hard_full/gch-stu004-20261116-r1-graph/launch/20261117/GCH-STU-005-full-evidence.json
ADDED
|
@@ -0,0 +1,95 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"work_id": "GCH-STU-005",
|
| 3 |
+
"issue_url": "https://github.com/dinhgia2106/NextKey_2026/issues/217",
|
| 4 |
+
"status": "running",
|
| 5 |
+
"run_id": "gch-stu004-20261116-r1-graph",
|
| 6 |
+
"identity": {
|
| 7 |
+
"git_commit": "520e09b7aac74c42b2c0441d5907781115578f4c",
|
| 8 |
+
"dirty_state": false,
|
| 9 |
+
"config_fingerprint": "198839d24ad833bdc41a9f8315f509862b6d23f33afd9d3e8ac20b0a94660ffe",
|
| 10 |
+
"dataset_fingerprint": "ccb6c5b7f1bddec623cd94a9fc146a4d268f35eaab666459de6b911732d4a32b",
|
| 11 |
+
"seed": 42,
|
| 12 |
+
"task_variant": "T2_JOINT",
|
| 13 |
+
"code_fingerprint": "da6f14d4b17ae156effa38ce70c3e929d07ad64598aa666db468b09a5120f529",
|
| 14 |
+
"dependency_fingerprint": "0f8fcfa7b5fb0386bcbc8f6c09868616cc00717f83484e41488d552947ef3a27",
|
| 15 |
+
"hardware": {
|
| 16 |
+
"device": "cuda",
|
| 17 |
+
"cuda_available": true,
|
| 18 |
+
"cuda_version": "13.0",
|
| 19 |
+
"cuda_device_name": "NVIDIA RTX PRO 6000 Blackwell Server Edition"
|
| 20 |
+
},
|
| 21 |
+
"precision": {
|
| 22 |
+
"float32_matmul": "highest",
|
| 23 |
+
"cuda_matmul_tf32": false,
|
| 24 |
+
"cudnn_tf32": true,
|
| 25 |
+
"mha_fastpath": true
|
| 26 |
+
},
|
| 27 |
+
"torch": "2.11.0+cu130",
|
| 28 |
+
"numpy": "2.2.6",
|
| 29 |
+
"python": "3.13.11 (main, Feb 3 2026, 03:08:54) [GCC 14.2.0]"
|
| 30 |
+
},
|
| 31 |
+
"notebook_url": "https://sb-a6a2e360c6bfa4d4.sb.molab.run/",
|
| 32 |
+
"source_checkpoint": "checkpoint_step_040000.pt",
|
| 33 |
+
"source_checkpoint_sha256": "a4c284f418e514cefc7a5a9d1772049d7592bcd253ee7a18a5efbfcbd1bcd5ac",
|
| 34 |
+
"source_checkpoint_revision": "44bf1b1a10151d016e2fd3fc8352ce7f5e0f932a",
|
| 35 |
+
"decision_publication": {
|
| 36 |
+
"revision": "9f8dc0cea05fb2e8eb2728e84b918a186f58f907",
|
| 37 |
+
"prefix": "student_screening/gch-stu004-20261116-r1/decision/20261117",
|
| 38 |
+
"winner": "graph",
|
| 39 |
+
"screening_snapshot": "44bf1b1a10151d016e2fd3fc8352ce7f5e0f932a",
|
| 40 |
+
"redownload_verified": true
|
| 41 |
+
},
|
| 42 |
+
"initial_step": 40000,
|
| 43 |
+
"initial_stream_offset": 40960000,
|
| 44 |
+
"full_horizon": 359851,
|
| 45 |
+
"parameter_count": 2078648,
|
| 46 |
+
"environment_fingerprint_matches": true,
|
| 47 |
+
"trained_checkpoint_continuation": {
|
| 48 |
+
"status": "passed",
|
| 49 |
+
"model_bitwise_equal": true,
|
| 50 |
+
"optimizer_bitwise_equal": true,
|
| 51 |
+
"scheduler_equal": true,
|
| 52 |
+
"rng_equal": true,
|
| 53 |
+
"metrics_equal": true,
|
| 54 |
+
"diagnostic_updates_per_branch": 2,
|
| 55 |
+
"metrics": [
|
| 56 |
+
{
|
| 57 |
+
"loss": 0.22505488991737366,
|
| 58 |
+
"grad_norm": 0.1361272782087326,
|
| 59 |
+
"lr": 0.0002928629560936078
|
| 60 |
+
},
|
| 61 |
+
{
|
| 62 |
+
"loss": 0.25460344552993774,
|
| 63 |
+
"grad_norm": 0.16958831250667572,
|
| 64 |
+
"lr": 0.0002928625629402227
|
| 65 |
+
}
|
| 66 |
+
],
|
| 67 |
+
"initial_step": 40000,
|
| 68 |
+
"stream_offsets": [
|
| 69 |
+
40961024,
|
| 70 |
+
40962048
|
| 71 |
+
],
|
| 72 |
+
"scope": "isolated copies; does not resume/extend pilot or initialize full training"
|
| 73 |
+
},
|
| 74 |
+
"gate_hf_revision": "b4fe5c9ec3d505183fce8c68e07543f4464cb643",
|
| 75 |
+
"observed_progress": {
|
| 76 |
+
"phase": "training",
|
| 77 |
+
"step": 40275,
|
| 78 |
+
"raw_records_consumed": 41241600,
|
| 79 |
+
"session_elapsed_seconds": 40.95349290099966,
|
| 80 |
+
"training_update_seconds": 3958.7811266599474
|
| 81 |
+
},
|
| 82 |
+
"observed_unix": 1794892005.4406335,
|
| 83 |
+
"observed_full_epoch_update_fraction": 0.11192132299201614,
|
| 84 |
+
"checkpoint_selection": "unchanged frozen100K selection EM@1, maximum; no test labels",
|
| 85 |
+
"official_final_validation": "unchanged Teacher pipeline after full epoch; not yet run",
|
| 86 |
+
"monitoring": {
|
| 87 |
+
"ai_interval_seconds": 1800,
|
| 88 |
+
"non_ai_keepalive_seconds": 300
|
| 89 |
+
},
|
| 90 |
+
"vm_segment": "r2",
|
| 91 |
+
"vm_replacements_so_far": 1,
|
| 92 |
+
"sealed_test_accessed": false,
|
| 93 |
+
"kd_or_knowledge_used": false,
|
| 94 |
+
"next_action": "Continue original Graph run to359851; preserve checkpoint/evidence sync and strict recovery if runtime fails."
|
| 95 |
+
}
|
student_hard_full/gch-stu004-20261116-r1-graph/launch/20261117/GCH-STU-005-full-plan.md
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# GCH-STU-005 — selected Graph HARD full-epoch continuation
|
| 2 |
+
|
| 3 |
+
Primary Codex at operator request. Receives completed GCH-STU-004 / #214,
|
| 4 |
+
PR215/216 and the operator plan's conditional40K→Full authorization. Scientific
|
| 5 |
+
source remains `520e09b7aac74c42b2c0441d5907781115578f4c`; no code/config/data/metric
|
| 6 |
+
change is introduced. Complete screening is frozen at HF
|
| 7 |
+
`44bf1b1a10151d016e2fd3fc8352ce7f5e0f932a` before any Full continuation.
|
| 8 |
+
|
| 9 |
+
Question/gate: can the selected Graph Student HARD continue the remaining
|
| 10 |
+
single epoch with exact state continuity? Graph leads selection EM at every
|
| 11 |
+
recorded10K interval, has the lowest final gold loss and all four final head
|
| 12 |
+
losses, stable finite training and under0.8% parameter spread. This is a
|
| 13 |
+
single-seed architecture operating decision, not a headline superiority claim.
|
| 14 |
+
|
| 15 |
+
Reuse run `gch-stu004-20261116-r1-graph`, original config
|
| 16 |
+
`configs/experiment/student_screen_graph_v1.yaml`, seed42, T2_JOINT, Gold-only
|
| 17 |
+
Teacher pipeline. Restore `checkpoint_step_040000.pt`, cursor40960000,
|
| 18 |
+
model/optimizer/full359851-step scheduler/all RNGs; omit the execution stop to
|
| 19 |
+
continue the existing epoch. Never initialize from weights alone or reset LR.
|
| 20 |
+
The scientific run/config identity is deliberately unchanged across work items.
|
| 21 |
+
|
| 22 |
+
Scope: runtime orchestration and docs/research map only. Before launch verify
|
| 23 |
+
source Git tree/commit/clean state, checkpoint SHA, full identity and a trained
|
| 24 |
+
checkpoint two-update CUDA bitwise continuation. VM-r2's entire dependency,
|
| 25 |
+
Python, hardware, Torch/CUDA and precision fingerprints already match VM-r1.
|
| 26 |
+
No concurrent candidate training. Preserve all original Graph40K evidence and
|
| 27 |
+
prior log segment. Stream only frozen TRAIN and use the unchanged Teacher
|
| 28 |
+
validation protocol and fixed sample SHA. No test, KD or Knowledge access.
|
| 29 |
+
|
| 30 |
+
Acceptance:359851 optimizer updates and368486843 raw TRAIN rows; unchanged
|
| 31 |
+
checkpoint selection/final Teacher validation; failure/status and resource logs;
|
| 32 |
+
HF checkpoint/evidence persistence; actual results and error analysis recorded.
|
| 33 |
+
AI checks every30 minutes (manual checks when requested),5-minute keepalive
|
| 34 |
+
without AI. On unavailable/replaced runtime, require fresh operator connection,
|
| 35 |
+
verify exact identities and resume the latest valid checkpoint rather than zero.
|
| 36 |
+
|
| 37 |
+
Expected artifact: reproducible Full HARD selected/last checkpoints and report.
|
| 38 |
+
Downstream: full HARD baseline and error analysis inform a separately recorded
|
| 39 |
+
Student KD program. Inherited ADR-029 source-audit limits and independent review
|
| 40 |
+
remain pending. No merge or headline/multi-seed claim is authorized by this plan.
|
student_hard_full/gch-stu004-20261116-r1-graph/launch/20261117/GCH-STU-005-full-status.md
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# GCH-STU-005 — Graph HARD Full continuation status
|
| 2 |
+
|
| 3 |
+
**Running, not completed.** Primary Codex at operator request; work
|
| 4 |
+
[#217](https://github.com/dinhgia2106/NextKey_2026/issues/217). Receives completed
|
| 5 |
+
[GCH-STU-004 screening/decision](GCH-STU-004-session-report.md) and the
|
| 6 |
+
[before-Full plan](GCH-STU-005-full-plan.md). Independent review remains pending.
|
| 7 |
+
|
| 8 |
+
Original run `gch-stu004-20261116-r1-graph`, T2_JOINT, seed42, scientific commit
|
| 9 |
+
`520e09b7aac74c42b2c0441d5907781115578f4c`, unchanged config/data/model/loss/decoder.
|
| 10 |
+
Full resumes the original40K checkpoint at raw cursor40960000 under the same
|
| 11 |
+
359851-update scheduler. Model/optimizer/scheduler/all RNGs are restored; no
|
| 12 |
+
weights-only restart, new run identity or LR reset. Checkpoint SHA and all
|
| 13 |
+
Python/Torch/CUDA/hardware/precision/dependency identities match.
|
| 14 |
+
|
| 15 |
+
The actual Graph40K trained-checkpoint two-update audit passes bitwise model,
|
| 16 |
+
optimizer, scheduler, RNG and metrics parity between continuous and save/load
|
| 17 |
+
branches. The diagnostic uses isolated copies; production starts from40000,
|
| 18 |
+
not40002. [Gate receipt](https://huggingface.co/datasets/Cuong2004/NextKey/blob/b4fe5c9ec3d505183fce8c68e07543f4464cb643/runs/gch-stu004-20261116-r1-graph/metrics/full-continuation-preflight.json).
|
| 19 |
+
|
| 20 |
+
At the launch verification, production reached40275/359851 updates and raw cursor
|
| 21 |
+
41241600 (11.1921% of update horizon). This is an observed launch status, not a
|
| 22 |
+
current live counter or completed epoch. Full has no new selection checkpoint
|
| 23 |
+
at this observation; the frozen40K best remains24.910%. Checkpoints/loss/selection
|
| 24 |
+
continue at the unchanged10000-update cadence. Full official validation has not
|
| 25 |
+
run. [Machine identity/gate/progress evidence](GCH-STU-005-full-evidence.json).
|
| 26 |
+
|
| 27 |
+
The screening decision/report was frozen before Full updates at HF revision
|
| 28 |
+
`9f8dc0cea05fb2e8eb2728e84b918a186f58f907`; original complete screening artifacts
|
| 29 |
+
remain pinned at `44bf1b1a10151d016e2fd3fc8352ce7f5e0f932a`. Do not relabel
|
| 30 |
+
later Full checkpoints as40K screening results.
|
| 31 |
+
|
| 32 |
+
[VM-r2](https://sb-a6a2e360c6bfa4d4.sb.molab.run/) is used. AI checks every30
|
| 33 |
+
minutes and manual checks on request;5-minute keepalive and checkpoint/evidence
|
| 34 |
+
publication use no AI. On runtime loss, require fresh operator connection and
|
| 35 |
+
strict checkpoint identity verification. One prior VM replacement is recorded.
|
| 36 |
+
|
| 37 |
+
Acceptance remains359851 updates/368486843 raw TRAIN rows, permitted unchanged
|
| 38 |
+
Teacher final validation, saved predictions/metrics, selected/last checkpoints,
|
| 39 |
+
error analysis and reproducible completion/recovery evidence. No completed HARD
|
| 40 |
+
baseline, KD/Knowledge use, sealed test, headline superiority or merge is claimed.
|
| 41 |
+
ADR-029 candidate source-audit limits remain. Downstream is the completed HARD
|
| 42 |
+
baseline/error analysis and a separately defined KD program.
|