CHORD Qwen3.5-2B student
The distilled Qwen3.5-2B encoder of CHORD, the coherence metric from the paper
Coherence-Aware Distributional Evaluation of Open-Ended Text Generation: Qwen3.5-2B with
LoRA adapters (merged into these weights) and a trained linear projection head
projector.pt (P_S: 2048 -> 256) on the last-token state of the coherence
prompt, trained to regress the Qwen3.5-27B teacher's top-256 PCA coordinates
with a single per-sample relative-MSE loss.
# pip install chord-metric[hf]
from chord import ChordScorer
scorer = ChordScorer("qwen3.5-2b-student") # loads this repository and its projection head
result = scorer.score(generated_texts, reference_texts)
Protocol: PromptEOL coherence template, max_length: 532 (512-token body plus
the reserved prompt suffix), last hidden state, whitespace text normalization,
RBF-MMD with a median bandwidth fitted on the reference pool. Raw MMD values are
comparable only within one encoder and protocol.
Files: model.safetensors (merged weights, float32), projector.pt (projection head),
bridge.json (projection-head metadata), tokenizer files. Training recipe:
distillation/configs/student_qwen3.5-2b/ in the code repository; training data and
teacher targets: mikezhu/chord-distill-data.
License
The weights are released under Apache-2.0, the license of the Qwen3.5 base model. The training corpus includes text sampled from OPT-1.3B, whose OPT-175B license does not allow data produced by OPT to be used for commercial or production purposes; check those terms before any commercial use of this model. The other training sources and their licenses are listed in the dataset card.
Citation
@misc{liu2026coherenceawaredistributionalevaluationopenended,
title={Coherence-Aware Distributional Evaluation of Open-Ended Text Generation},
author={Jinnuo Liu and Junhao Zhu and Weifeng Jiang and Haoming Liu and Hongyi Wen},
year={2026},
eprint={2609.34240},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2609.34240},
}