Jev-LCT-Qwen2.5-1.5B
Qwen2.5-1.5B with its top two layers fine-tuned to loop up to four times; it reads a state and returns a probability per option for choice, noul and score questions, taking confidence from how the loops converge. Runs on the author's own inference code.
A personal research release by Cao Haowei. Only the top 2 of 28 layers and a small pointer head are fine-tuned, on a small training mixture; noul is P(yes) over yes/no tokens and score is the expected value over ordered levels. An adaptive early exit runs 1 to 4 loops. On the maker's own 300-question suite, accuracy equals its single-pass baseline; the claimed gain is calibration and latency, not independently checked. It needs the author's standalone loader plus a pickled auxiliary head; loading it as plain Qwen2 drops the loops. The repo ships a server shaped like Jev's /v1/systemone. Siblings: Jev-LCT-Qwen2.5-0.5B, Jev-LCT-Qwen3-8B (81.7% in the same run) and an adapters repo.
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
At a glance
| Parameters | 1.54B |
| Base model | Qwen/Qwen2.5-1.5B |
| Maker | Cao Haowei |
| Released | 2026-09-25 |
| License | apache-2.0 |
| Reported accuracy | 70.3% |
| Reported latency | 61.9 ms average per question on an RTX 3090 Ti |
Get the weights
pip install systemonemodels
systemone pull cao-haowei/jev-lct
The files are served from the maker's Hugging Face repository, CaoHaoWei/Jev-LCT-Qwen2.5-1.5B, and verified against the checksums recorded here.