Beats fine-tuning, with no training run.
Duplex made a frozen model behave as if it had been fine-tuned, and beat the best adapter from a tuned sweep of LoRA, QLoRA and DoRA in all six tests: three public intent datasets on two open model families, three runs each.
- 6 / 6Tests where Duplex beat a fine-tuned adapter+30.3 points on average
- 0Retraining to move to another model familyThe adapters moved in 0 of 6
- 60 to 79%Answered from what it learnedOn either model family
A frozen model, adapted without touching it.
Step 1 of 5: Frozen
Ahead in all six tests.
Accuracy on BANKING77, CLINC150 and HWU64, on Qwen2.5 and SmolLM2.
Step 1 of 3: Frozen
On their own, the frozen models scored between 11% and 49% across the six tests.
It keeps learning on the job.
As verified examples arrive, Duplex answers more itself, and stays right.
Step 1 of 3: 40 examples
After 40 verified examples, Duplex answered 1.7% of requests itself, all of them right.
Where Duplex goes further than fine-tuning.
General ability, removal, moving models and repeat requests, side by side.
Step 1 of 4: General ability
On a general-knowledge check, Duplex left the model unchanged in all six tests. The adapter cost it 1 to 4 points.
How it was run.
- The task
- Intent classification: read a message and return exactly one intent label.
- Datasets
- BANKING77 (77 intents), CLINC150 (151) and HWU64 (64), all public.
- Models
- Qwen2.5-1.5b-instruct and SmolLM2-1.7b-instruct, frozen.
- Scale
- 300 training examples per run, three runs for each dataset and model family: 18 runs.
- Compared with
- The frozen model on its own, and the best adapter from a tuned sweep of LoRA, QLoRA and DoRA.
- Kept equal
- The same frozen model, data, splits, prompt and answer parsing for every approach.
What it means for your work.
Adaptation without a training run.
Duplex beat the tuned adapter in every test, by 30.3 points on average.
Your model stays your model.
Its general knowledge was unchanged, and any single thing Duplex learned can be removed exactly.
Switch models freely.
What Duplex learned moved to the other model family with no retraining.
See it on your own work.
From your own assistant to every agent your company runs, a pilot puts Duplex on your own work, with the measures agreed before it begins.