Sampler checkpoint
Use this identity for base-versus-trained inference. The concrete remote identifier stays private.
The adapter exists. That is not the same as improvement.
One tiny run proved the plumbing. Two complete passes trained the real cohort. The checkpoints are usable, but checkpoint existence is an engineering result—not a quality verdict.
The proof run, two full passes, remote checkpoints, failures, and recoveries.
01 / Proof run
Start with the smallest billable truth
Before committing the full cohort, one reviewed row went through the complete supervised fine-tuning path. It rendered into model tokens, completed one optimizer step, and saved both serving and resumable state.
Reviewed rows1Smoke-test only
Optimizer steps1Completed
Output2Checkpoint types
ClaimPipeNot improvement
The result removed uncertainty about authentication, token rendering, loss computation, and checkpoint saving. It did not contain enough examples to say anything serious about behavior.
02 / Full run
Two rank-32 passes over 42 accepted rows
The real cohort trained as two bounded passes. Each pass used all 42 accepted rows and completed 42 optimizer steps. The second pass continued the first adapter rather than starting an unrelated model.
Input42 accepted rows
31 text-only plus 11 audio-conditioned.
Pass one42 optimizer steps
A rank-32 adapter learns the first update.
Pass two42 optimizer steps
The same adapter receives a second complete pass.
ResultFinal checkpoints
Serving weights and resumable state are saved remotely.
In total, the two complete passes produced 84 optimizer steps. The run used Tinker for hosted training. Modal credits remained unused because no separate GPU deployment was required.
03 / Checkpoints
Serve and resume are not the same artifact
The final training result has two operational forms. A sampler checkpoint combines the adapter with inference-time loading so the model can answer prompts. A training-state checkpoint preserves optimizer state so an intentional continuation can resume correctly.
Training state
Use this only when additional training should continue the same adapter and optimizer history.
Where is the LoRA?
It is stored remotely by the training service. The public website contains neither its weights nor a private loading URI.
04 / Incident log
The first full attempt failed before step zero
An early attempt hung on macOS and ended in a connection error before the optimizer advanced. That distinction mattered: there was no partial model to celebrate or discard. The failed attempt was an infrastructure incident, not a training result.
Connection failure
The request failed before step zero. Logs were preserved and no checkpoint was claimed.
SDK certificate issue
The provider's published status guidance pointed to a corrected SDK version.
Same data, corrected environment
The dataset and bounded plan remained fixed while the transport dependency changed.
Both passes completed
Every planned step ran and final artifacts were saved.
The correct claim
Training completed and produced a usable adapter. That does not prove the model improved; evaluation must answer that separately.