Contrastive-LM / CLM

Contrastive Language Models
 A System One Model for Fast and Generalizable Decision-Making 
| 📄 Blog | 🗣️ Discord | 🤗 Data & Models | 📚 API Reference | 🛠️ Fine-Tuning Tutorial |
🔥 **Contrastive Language Models (CLMs)** are a new class of **System One
model** trained with a **contrastive learning** objective that connects
**states and actions**. This repo serves **CLM-8B** behind a
TypeSafe-compatible API.
- **CLM-8B** is pre-trained on **60M Nemotron Q&A pairs**, mid-trained on
 **30M synthetic hard negatives**, and post-trained on **1M agentic
 trajectories**.
- It performs on par with **Jev** across computer-use, gaming and tool-calling
 tasks with up to **9× lower latency**. With lightweight fine-tuning it sets a
 new SOTA as a verifier on agentic coding benchmarks: **Terminal-Bench 2.1
 (87.6%)** and **DeepSWE (81.6%)**.
- **States and actions are disaggregated**, so their embeddings are cached and
 reused independently, which makes training and serving cheap and blazing fast!
We invite the community to plug it into their own agents and benchmarks!
---
## Installation
 [code] 
To install the latest from a clone:
 [code] 
---
## Quickstart
### Serve
 [code] 
States longer than 2048 tokens are truncated. For longer states, raise both limits
together, e.g. `--max-model-len 8192` on `vllm serve` and `clm-serve --max-tokens 8192`
(needs more GPU memory).
### Ask typed questions about a state
 [code] 
Questions may be `Noul` / `Choice` / `Score` objects or plain wire-format
dicts, so a request written for TypeSafe replays as
`client.system_one(state, questions)`.
### Rank candidates directly
`system_one` is built on one primitive: score a candidate against a state.
For free-form candidates (best-of-N answers, tool names, next moves) use the
in-process engine's `rank`:
 [code] 
---
## Playground
`clm-serve` also serves a web UI at `/` (`http://localhost:8700/` by default).
Write a state, add typed questions, and see CLM's answer distributions; every
request is also shown as JSON, `curl` and Python. A **Rank** tab ranks any
candidate set, and links are shareable.
 Captured against a real clm-serve ( clm-latest , Qwen3-8B encoder on one RTX 4090). 
Remote server? `ssh -L 8700:localhost:8700 `. API only: `clm-serve --no-ui`.
---
## Results
### Zero-shot evaluation
Across **computer-use, gaming and tool-calling tasks**, CLM-8B performs on par
with Jev while running **up to 9× faster**. The speedups are largest when the
number of candidate actions is large (WikiRacing) or when actions are reused
across states (the T-Rex game). The T-Rex benchmark ships in this repo:
see examples/t_rex.
### Agentic benchmarks: CLM as a verifier
For each task we sample several candidate solutions (**Opus 5** for DeepSWE,
**Fable 5** for Terminal-Bench 2.1), and CLM or Jev acts as the verifier that
picks the best one. Evaluated on **38 held-out DeepSWE tasks** and **30
held-out Terminal-Bench 2.1 tasks**; latency on an H100. Jev fails to serve as
a verifier for these long-h