Categories / huggingface/transformers / evidence
Transformers centralizes model definitions for text, vision, audio and multimodal models so training and inference frameworks across the ecosystem can share a single implementation.
This audit's judge read a single 13.6k-token evidence dossier, compiled mechanically under per-section budgets. Below: each section, what went in, and what was cut. Citations on the audit page resolve here.
| # | Section / contents | Mode | Spent / budget |
|---|---|---|---|
| [28] | Vitals + README (distilled) | SECTION | 1.4k / 1.5k |
| [1] | README.md · distilled: badges/TOC stripped | DOSSIER | 1.4k |
| [2] | Structure (dir-aggregated) | SECTION | 738 / 1.0k |
| [29] | Health + Static digest | SECTION | 1.3k / 2.0k |
| [3] | (github api) | PROBE | 625 |
| [4] | (static analysis) | PROBE | 625 |
| [30] | Manifests + lockfiles + CI (verbatim) | SECTION | 1.2k / 1.2k |
| [5] | pyproject.toml | DOSSIER | 294 |
| [6] | .github/workflows/dependents-smoke-test.yml | DOSSIER | 294 |
| [7] | .github/workflows/self-scheduled-amd-mi250-caller.yml | DOSSIER | 294 |
| [8] | .github/workflows/release-conda.yml | DOSSIER | 294 |
| [31] | Core files (complete / whole-function) | SECTION | 5.0k / 5.0k |
| [9] | src/transformers/modeling_utils.py · partial: 1/132 whole functions | DOSSIER | 2.5k |
| [10] | src/transformers/testing_utils.py · partial: 1/297 whole functions | DOSSIER | 2.1k |
| [11] | src/transformers/image_processing_utils.py · partial: 1/21 whole functions | DOSSIER | 310 |
| cut: src/transformers/configuration_utils.py — over budget (17507 tok), functions too large to extract | |||
| cut: src/transformers/training_args.py — over budget (35405 tok), functions too large to extract | |||
| cut: src/transformers/processing_utils.py — over budget (29581 tok), functions too large to extract | |||
| cut: src/transformers/modeling_outputs.py — over budget (26422 tok), functions too large to extract | |||
| [32] | Test reality | SECTION | 1.4k / 1.5k |
| [12] | tests/models/markuplm/test_feature_extraction_markuplm.py | DOSSIER | 1.4k |
| [33] | Probe answers | SECTION | 1.4k / 1.5k |
| [13] probe-deps[14] probe-license[15] probe-ci[16] probe-todo[17] probe-assertions[18] probe-top-tested[19] probe-cve[20] probe-claim-1[21] probe-claim-2[22] probe-claim-3[23] probe-claim-4[24] probe-claim-5[25] probe-claim-6 | |||
| [26] | Outlines (files not read above) | SECTION | 323 / 500 |
| [34] | Coverage notes (what was cut & why) | SECTION | 257 / 400 |
| [27] | (composer) | DOSSIER | 257 |
| 6320 files mapped · 1 complete, 3 whole-function · 13 probes | 13.6k / 15k cap | ||
Raw signals behind the scores. Source tags: COMPUTED derived from the clone · API GitHub/registry data · JUDGE model estimate
Every judgment on the audit page, mapped to what backs it.
| Claim | Backed by | Kind |
|---|---|---|
| pipeline() quickstart runs a working text-generation call after pip install | vitals-readme | OBSERVATION |
| Dependency scan found 17 critical and 83 high-severity CVEs with no lockfile | probe-cve probe-deps | MEASUREMENT |
| 85% of open issues are stale beyond 180 days | health-static | MEASUREMENT |
| 36.2% of the codebase is duplicated, concentrated in high-churn hotspot files | static | MEASUREMENT |
How much of the repository this audit actually covered, who probably wrote it, and how heavy it is — informational, never part of the grade.
structure mapped: 6320 files · dossier: 13562 tok · complete files: 0 · whole-function: 3 · probes: 13
large long-running multi-contributor project with human commit history and governance files