BioDecision-4B v2
A 4B biomedical decision model by Dimitris Papadopoulos: Qwen3.5-4B tuned with Together AI's Tev1 recipe on 1.75M pharma, clinical-trial and literature decisions. One forward pass gives a temperature-scaled probability per option. Research use only.
Jev-inspired: trained with Together AI's Tev1 recipe to put its answer on a single option letter. Takes TypeSafe's systemone request format (choice with 2 to 24 options, yes/no and score) and ships an example server. Temperatures are fitted per task family; no calibration error is reported for v2. In the maker's own tests it scores 75.3% on 63,432 overlap-screened held-out decisions, against 69.4% for v1. Answers change on 6.9% of decisions when options are reordered; about 8.5% of training rows contain LLM-written material. Research use only because several training sources are non-commercial, so it is not mirrored or packaged here. Not validated for patient care.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
- classify — assigns a category from a fixed taxonomy
At a glance
| Parameters | 4B |
| Base model | qwen/qwen3.5-4b |
| Maker | Dimitris Papadopoulos |
| Released | 2026-09-28 |
| License | Research use only (several training sources are non-commercial); not validated for patient care |
| Reported accuracy | 75.3% |
| Reported latency | 39 ms median per decision with vLLM on one A100 80GB, bf16 |
Get the weights
pip install systemonemodels
systemone pull lighteternal/biodecision
The files are served from the maker's Hugging Face repository, lighteternal/biodecision-v2-4b, and verified against the checksums recorded here.