bulbface/FunDM-caches

FunDM paper caches

Cached numerical features and teacher scores for reproducing FunDM. All selected feature tensors use Qwen3-8B (hidden width 4096). No 27B artifacts are included.

Contents

• feats_kd/: structured-prompt MCQ states and teacher scores at layers 16, 20, 24, 28, and 32, including the original and expanded training pools.
• feats/: supervised CE and encoding-ablation features and an opaque group map.
• feats_cf/: whole-run CoderForge states, targets, and task groups.
• data/clm_emb/: separate CoderForge/DeepSWE embeddings and CoderForge joint features.
• data/clm_choice/embeddings/: CLM's hash-keyed MCQ embedding cache.
• results/coderforge/judge.jsonl: numerical teacher judgments.
• results/clm_mmlu_pro_preds.jsonl: CLM reference probabilities for serving validation.
• feats_kd_oneline/train_L32.pt: the legacy prompt-comparison cache.

This is a whitelist of paper artifacts, not a copy of the research working directory. No backbone weights, trained head checkpoints, raw benchmark questions/options, raw trajectories, smoke tests, or unused experimental caches are bundled.

Download

Install the FunDM environment and clone its CLM submodule first. From a fresh FunDM repository root:

import json
from pathlib import Path
from huggingface_hub import HfApi, hf_hub_download, snapshot_download

repo = "bulbface/FunDM-caches"
revision = HfApi().dataset_info(repo).sha
manifest_path = hf_hub_download(repo, "manifest.json", repo_type="dataset", revision=revision)
manifest = json.loads(Path(manifest_path).read_text())
snapshot_download(repo, repo_type="dataset", revision=revision, local_dir=".",
                  allow_patterns=[item["path"] for item in manifest["files"]])

Use the pinned dataset revision and checksum-verification instructions in the FunDM reproduction guide. The allow-list avoids overwriting the code repository's README and license. manifest.json records each artifact's path, byte size, and SHA-256 checksum.

Caches avoid repeated feature extraction. The table commands still fit heads and run any missing baseline evaluations. Tables 2 and 12 check archived timing reports unless a fresh report is supplied. Upstream datasets and model/head weights are obtained from their original hosts when needed.

Metadata and provenance

The historical structured cache label old-structured is preserved. These caches are intended for the paper's kd_train and table workflows. Do not relabel their prompt metadata or assume a cache-trained checkpoint is compatible with the public raw predict command, which expects metadata from its extract workflow.

The original group map stored normalized question text. Its replacement uses ordered opaque IDs and also covers extra-pool fallback IDs. Every training and validation index is unchanged in the 25 comparisons recorded in group-encoding-check.json. Numerical tensors and targets are unchanged.

Public task/run IDs retain GitHub owner/repository handles for reproducibility. Upstream data may contain bias or objectionable material; numerical caches are not a privacy guarantee. The features cover English MCQ benchmarks and coding-agent trajectories as described in the paper; they do not document demographic coverage.

Terms

This collection has mixed source terms, not a blanket Apache-2.0 license. SciQ/RACE-derived mixed training caches are supplied for non-commercial research. See LICENSES.md for attribution, restrictions, and source links, and provenance.json for pinned revisions. The same license file includes source licenses for all 8,784 selected CoderForge runs. Original authors retain their rights.