Pilot-0.3 — continued training

This is not a third independent model birth. It is continued training of Pilot-0.2 / Local-70M-0 — the same 69,970,848-parameter model, resumed from that model’s final checkpoint and trained to 1,000,013,824 cumulative tokens.

What this run is

This is not a third independent model birth

It is continued training of Pilot-0.2 / Local-70M-0 — the same 69,970,848-parameter model, started from that model’s final checkpoint.

There is no new initialization and no new seed. Counting this as a third birth would turn two independent observations into three.

Parent model
Pilot-0.2 / Local-70M-0
Parameters
69,970,848 identical to the parent — same model
Cumulative tokens
1,000,013,824
Cumulative steps
61,036
Final validation CE
2.557595558464527

Explicit confounds

This is not a controlled token-only intervention. The schedule restart is one of the changed conditions, so Pilot-0.2 and Pilot-0.3 cannot be compared as two different births.

The run includes a second exposure to the same corpus and a restarted learning-rate schedule. Any observed change therefore has at least two candidate causes, and this experiment cannot separate them.

Validation curve

Evidence authority: Diagnostic-only

Pilot-0.3 validation cross-entropy across continued training Validation cross-entropy from the Pilot-0.2 source final checkpoint through continued training to one billion cumulative tokens. The curve rises first, then falls below the source level. A vertical marker shows where the learning-rate schedule restarted. The marker records when the restart happened; it does not assert that the restart caused the change, because there is no control arm.CE2.552.602.652.70schedule restart occurred herePilot-0.2 source final — cumulative step 32,470, validation CE 2.62219978121% — cumulative step 32,756, validation CE 2.68356869375% — cumulative step 33,899, validation CE 2.708705712110% — cumulative step 35,327, validation CE 2.706871256225% — cumulative step 39,612, validation CE 2.675921417850% — cumulative step 46,754, validation CE 2.622445117775% — cumulative step 53,895, validation CE 2.5734781809100% — cumulative step 61,036, validation CE 2.557595558532,47025%50%75%100%cumulative training step → 61,036
Validation cross-entropy across continued training. CE rises at the start of the continuation and later falls past the source level to 2.55760. The marker records where the schedule restarted — it is a timing marker, not a causal claim. Source §7.2.
Pilot-0.3 validation curve — source §7.2, diagnostic-only.
PointCumulative stepValidation CE
Pilot-0.2 source final32,4702.6221997812
1%32,7562.6835686937
5%33,8992.7087057121
10%35,3272.7068712562
25%39,6122.6759214178
50%46,7542.6224451177
75%53,8952.5734781809
100%61,0362.5575955585

The validation regression in the first section happened at the same time as the schedule restart. CE later fell past the source level to 2.55760. Because there is no control arm, the cause cannot be attributed to the schedule alone or to the second corpus pass alone.

Measured profile shift

Evidence authority: Diagnostic-only

In plain language

Training the same model for longer made its general language score better — and made two of the measured signals worse.

The two measures that had grown most during the original training (I and A) both declined across the continuation, while an easy measure (R) rose. So “train longer, get better at everything” did not hold here.

This is one continuation of one model, measured on the same questions. It is an example, not a rule about training in general.

Technical evidence

Matched items, n=128 per cell, at the 25%, 50% and 75% checkpoints of the continuation.

  • I@8: +0.452 → +0.371 → +0.315
  • A@0: +0.325 → +0.283 → +0.267
  • R@1: +0.141 → +0.109 → +0.266
  • Validation CE: 2.6759 → 2.6224 → 2.5735

The general language-model validation metric improved while the measured I and A signals declined and easy R rose. This profile shift rests on a single continuation lineage, the same items, and n=128 diagnostic sampling.

Correct reading: this happened in this continuation lineage. Incorrect reading: continued training harms reasoning.

Pilot-0.3 measured phenotype across three checkpoints, with the final checkpoint unmeasured Normalized scores for eight cells at the twenty-five, fifty and seventy-five percent checkpoints of continued training, over n=128 matched items. The one hundred percent slot is drawn as an explicit labelled gap marked UNMEASURED. The final checkpoint and its receipt exist, but no capability battery was run on it. No line is extended into that region, no value is interpolated, and it is not treated as zero.0 = chance0.20.4M@64 at 25% — 41/128, normalized +0.094M@64 at 50% — 41/128, normalized +0.094M@64 at 75% — 43/128, normalized +0.115 R@1 at 25% — 73/128, normalized +0.141R@1 at 50% — 71/128, normalized +0.109R@1 at 75% — 81/128, normalized +0.266 I@8 at 25% — 60/128, normalized +0.452I@8 at 50% — 50/128, normalized +0.371I@8 at 75% — 43/128, normalized +0.315 C@True at 25% — 3/128, normalized −0.008C@True at 50% — 2/128, normalized −0.016C@True at 75% — 5/128, normalized +0.008 Q@1 at 25% — 5/128, normalized −0.024Q@1 at 50% — 4/128, normalized −0.032Q@1 at 75% — 6/128, normalized −0.016 V@8 at 25% — 20/128, normalized +0.036V@8 at 50% — 20/128, normalized +0.036V@8 at 75% — 20/128, normalized +0.036 A@0 at 25% — 47/128, normalized +0.325A@0 at 50% — 42/128, normalized +0.283A@0 at 75% — 40/128, normalized +0.267 T@modulus at 25% — 14/128, normalized +0.095T@modulus at 50% — 14/128, normalized +0.095T@modulus at 75% — 15/128, normalized +0.103100%UNMEASUREDM@64 +0.115 R@1 +0.266 I@8 +0.315 C@True +0.008 Q@1 −0.016 V@8 +0.036 A@0 +0.267 T@modulus +0.10325%50%75%100%
Measured cells at the 25%, 50% and 75% continuation checkpoints. The 100% slot is an explicit gap labelled UNMEASURED: the checkpoint exists and its receipt matches, but no capability battery was run on it. Nothing is drawn into that region. Source §7.3 measured, §8.5 unmeasured.
Pilot-0.3 measured checkpoints — source §7.3, diagnostic-only, n=128 matched items per cell. The 100% row is unmeasured, source §8.5.
CheckpointM@64R@1I@8C@TrueQ@1V@8A@0T@modulus
25%41/128
+0.094
73/128
+0.141
60/128
+0.452
3/128
−0.008
5/128
−0.024
20/128
+0.036
47/128
+0.325
14/128
+0.095
50%41/128
+0.094
71/128
+0.109
50/128
+0.371
2/128
−0.016
4/128
−0.032
20/128
+0.036
42/128
+0.283
14/128
+0.095
75%43/128
+0.115
81/128
+0.266
43/128
+0.315
5/128
+0.008
6/128
−0.016
20/128
+0.036
40/128
+0.267
15/128
+0.103
100%UNMEASURED — the final checkpoint and receipt exist, but no capability artifact was produced (source §8.5). Not zero, not a failure.

Final capability state

The 100% final checkpoint is UNMEASURED

The durable diagnostic job requested and completed only the 25%, 50% and 75% checkpoints. The 1%, 5% and 10% checkpoint files exist as training records but no capability battery was run on them. The 100% final checkpoint and its receipt exist, but there is no n=128 or n=800 capability artifact for it, and the durable records do not contain a sufficient reason for why the final scan was not run.

The 100% checkpoint therefore cannot be called a failure. The correct label is unmeasured. Filling the missing cell with a zero score, or presenting 75% as if it were the final state, is not permitted.

Unmeasured is not zero, not chance, not a failure, and not an interpolated value. The 75% checkpoint is not presented as the final state of this model.

Final checkpoint
Exists; receipt matches
Capability artifact
Does not exist
Correct label
UNMEASURED
What would close this gap
Running the same frozen n=128 diagnostic scan on the 100% checkpoint — see Limits.

Training reliability

Run status
COMPLETED
Final checkpoint size
915,599,240 bytes
Hash receipt
Final checkpoint hash matched the receipt.
Non-finite gradient events
13
Handling
The 13 non-finite gradient events were handled by same-batch GradScaler backoff and the run completed.

These events are visible in the fail-closed training records and were not hidden from the results.

These events are reported because they happened, and they are not inflated into a claim that the run was unstable. The run completed, and the final hash matched its receipt.

Source: Scientific Evidence Inventory v1.0 (evidence cutoff 30 August 2026), §7.1, §7.2, §7.3, §7.4, §8.5.