Next Experiment / live investigation

Investigate a fine-tuning failure

Upload your data and describe the failure. Follow Nemotron’s hypotheses and checks, then compare a proposed fix with a control trained through Token Factory.

Free testing · $20 shared budgetRecorded investigations
What a live investigation includes

One diagnostic pair, with two seeds per arm, using Qwen3-0.6B. Answers are scored by exact match with a 32-token output limit. The submitted recipe and losses guide the investigation; the original training run is not verified. Full retraining is available as an exported proposal.

Checking the live-run budget…

Describe the failure

Fixed settings: LoRA alpha = 2 × rank, batch size 4, context 8192, packing off.

Upload four distinct splits

Download an example with all four files, or use your own data.

Chat JSONL: one object per line, with a messages array and a final assistant answer. Use short reference answers (at most 100 characters). Each row must be below 4 KB. Identical records across splits are rejected.

See the row format
{"messages":[{"role":"user","content":"My parcel is late. Choose delivery or billing."},{"role":"assistant","content":"delivery"}]}

Results on these small samples are exploratory. Benchmark thresholds are not calibrated to your data; the result may be inconclusive.

Start the investigation

Your files are processed by this service and sent to Nebius for training. Probe summaries may include excerpts sent to Nemotron. Run evidence and reports are stored privately on Hugging Face; only this browser’s access key retrieves them through the app. Do not upload secrets or personal data. The service runs one investigation at a time.

Checks usually take minutes; training can take longer depending on the queue. Keep this tab open. Your access key is saved in this browser so you can reconnect after a refresh.