TRAINING AGENT
Fastino Training Agent researches your task, builds datasets, runs SFT and RL experiments, evaluates and deploys models, in one prompt. Model engineering in hours, not weeks.
coding agent
Fastino’s plan · nothing trains until you approve
base Nemotron 3.5 Lightning 30B-A3B
method SFT · rank-32 LoRA adapter
data 6 finance skills · held-out split
bar beat base on FinQA and TAT-QA
spend quoted range · your ceiling
› approve
• Called fastino (approve_plan)
├ data engine ✓ curated and audited
├ experiments ✓ SFT runs complete
└ evaluation ✓ champion clears bar
Report saved: training-report.md ↗
FinQA 15.9 → 59.2 +43.4
TAT-QA 19.0 → 56.6 +37.6
ConvFinQA 15.0 → 57.3 unseen
? Promote the champion to a live endpoint? needs your confirmation
Autoresearch agent
A closed-loop autoresearch agent
One prompt in, a model that beats your baseline out. No human in the loop between your plan approval and the go-live call.
01
Training plan
Researches your task and constraints, selects a base model, and proposes a training plan for your approval.
02
Data engine
Finds, curates, and generates training data. Builds a separate evaluation set before training begins.
03
Experiment
Runs SFT or RL experiments in parallel and uses the results to refine the data and training strategy.
04
Evaluation
Compares candidates against the baseline on held-out data, checks for regressions, and selects the strongest checkpoint.
05
Deployment
Delivers the selected checkpoint, training data, and a report explaining the experiments and results.
The agent documents its experiments, data sources, evaluation results, and contamination checks in one report your team can review.
agent-training-report.md
TRAINING REPORT / FINANCE
Fastino-Nemotron-3.5-Lightning-Finance
Base: Nemotron-3.5-Lightning-30B-A3B SFT LoRA · rank 32 · 2 epochs 6 skills, one adapter
Candidate A meets the success bar. Both pre-registered gates passed, and the gain transfers to a benchmark it never trained on.
59.2
FinQA accuracy
+43.4
points vs untuned base
2 / 2
gates passed
1
Evaluation gates
Gate
Base → A
Change
Verdict
FinQA gain
15.86 → 59.23
+43.37
Passed
TAT-QA gain (F1)
19.01 → 56.63
+37.62
Passed
ConvFinQA · not trained on
15.00 → 57.33
+42.33
Transfer
2
Candidates
A
Six specialists, mixed by gain retained
Recommended
B
Same mix, tool-use set not upweighted
Tool use lost
C
Smaller SEC-Num share
Extraction dropped
3
Data
13,698 rows built from six specialist winners. Hold-out sets fixed before training; transfer benchmarks never trained on. Selected on hold-out behavior, not training loss.
DEPLOYMENT STATUS / PREVIEW ONLY
Candidate A qualifies for delivery. Going live needs your confirmation.
Also in the report: error analysis, proposed next experiments, and an audit record of model, dataset and run IDs.
Example layout. Scores from the published Nemotron 3.5 Lightning Finance run.
Our Nemotron 3.5 Lightning specialists were post-trained end to end by the Training Agent. Each one runs on 3B active parameters.
Request access
BUILT WITH THE AGENT
Nemotron 3.5 Lightning Finance
Six finance skills in one adapter: financial programs, table arithmetic and numerical extraction.
FinQA
59.2
+43.4
TAT-QA
56.6
+37.6
FinEntity
79.5
+19.4
ConvFinQA *
57.3
+42.3
Base
Fine-tuned
* not trained on
Case study ↗
BUILT WITH THE AGENT
Nemotron 3.5 Lightning Healthcare
Six clinical skills in one adapter: concept extraction, error flagging and biomedical reasoning.
BC5CDR
72.5
+24.6
MedMentions
40.0
+20.6
MEDEC
65.0
+11.3
BioRED *
41.8
+15.9
Base
Fine-tuned
* not trained on
Case study ↗
FAQ