CoolFace
Modelpublic

DefEki/qwen35-ple-auditable-ngram-memory

sourceHugging Faceapache-2.0updated 15d agoView on Hugging Face
0likes
Model Card

qwen35-ple Auditable N-Gram Memory Artifacts

This release contains the reproducible artifacts for the paper Auditable N-Gram Memory for Small Language Models.

Links

  • —Code repository: https://github.com/QingGo/qwen35-ple
  • —Paper source: https://github.com/QingGo/qwen35-ple/blob/main/paper/paper.typ
  • —Compiled PDF: https://github.com/QingGo/qwen35-ple/blob/main/paper.pdf
  • —Evaluation card: https://github.com/QingGo/qwen35-ple/blob/main/docs/evaluation-card-paper.md
  • —Reproducibility manifest: https://github.com/QingGo/qwen35-ple/blob/main/docs/reproducibility-manifest.md

Contents

  • —projectors/: PLE Projector checkpoints (10k training, seeds 0-4).
  • —adapters/: Purified OPSD MoRA adapters (seeds 0-2).
  • —datasets/: PLE projector local-continuation datasets (1k and 10k).
  • —configs/: N-gram fusion router / token policy / projector configs.
  • —scripts/: Evaluation, training, fairness/sensitivity scripts.
  • —results/: HumanEval, TriviaQA, pass@k, LLM judge, sensitivity, CPU benchmark results.
  • —docs/: Evaluation card, reproducibility manifest, evidence notes.
  • —artifact-manifest.json / SHA256SUMS: checksums and provenance.

Use

bash
# Reproduce PLE projector training (example)
python scripts/train_ple_projector.py   --model <Qwen3.5-0.8B>   --dataset datasets/ple-projector-dataset-10k.jsonl   --max-train-samples 7000 --max-eval-samples 300 --steps 100   --device cuda --seed 0 --output projector-seed0.json

License

Artifacts are released under Apache-2.0 unless noted otherwise. Model weights remain subject to the upstream Qwen license.