PKU-KCL-Meteor-2B
PKU-KCL-Meteor-2B is a decision model from the Knowledge Computing Lab (KCL), Peking University. Given a state and typed questions, it answers each question in one forward pass with a probability for every option, without generating text. It is built on Qwen/Qwen3.5-2B (2.21B parameters).
- Questions:
choice(up to 255 options),noul(a yes / no probability) andscore(2 to 10 ordered levels), in the System One request format. - One pass per question: the prompt lists every option under a single-token answer code, and a readout scores those codes at the last prompt position. No text is generated.
- No truncation: a question longer than the maximum model length (262,144 tokens) is refused, never cut.
Results
Decision Index 0.3 public suite (kit commit 25c6ec7b): index 60.95 (raw 70.98) over all 140,178 scored requests, 140,178 answered (0 unsupported or failed). Chance-corrected area scores: knowledge 37.1, language 62.3, retrieval 71.6, tools 84.2, arts 55.3.
Quick start
pip install -r serving/requirements.txt
pip install --no-build-isolation causal-conv1d==1.7.0
from meteor import Meteor
model = Meteor() # this directory, cuda:0
model.decide(
"The traffic light turned red as the car approached.",
{"action": {"type": "choice", "criteria": {"stop": "Stop", "go": "Keep going"}},
"safe": {"type": "noul"}},
)
# {"model": "PKU-KCL-Meteor-2B", "answers": {"action": {"type": "choice", "choice": ..., "probabilities": {...},
# "confidence": ...}, "safe": {"type": "noul", "noul": ...}}, "usage": {"input_tokens": ..., "output_tokens": 0}}
The Decision Index kit runs it in process with --engine decision_index_engine:MeteorEngine, and
serving/server.py serves /v1/systemone; see serving/SERVING.md.
Good to know
- Text only, on one NVIDIA GPU: BF16 weights with an FP32 readout.
- A request's questions run in batches, shortest first; answers do not depend on batching beyond floating-point rounding.
MODEL_MANIFEST.jsonlists every file's SHA-256; loading checks small files' hashes and large files' sizes.
License
Copyright 2026 Knowledge Computing Lab (KCL), Peking University.
- Weights: CC BY-NC 4.0: free for research and other non-commercial use, with attribution. For commercial use, contact the Knowledge Computing Lab (KCL), Peking University.
- Code (
meteor/,serving/,decision_index_engine.py): Apache License 2.0. - Built on Qwen/Qwen3.5-2B (Apache License 2.0, LICENSE-QWEN); see NOTICE.
- Downloads last month
- 88