fix(classifier): local_encoder default model is gated, switch to bart-large-mnli #48

alee · 2026-09-07T00:34:20Z

alee commented

2026-09-07 00:34:20 +00:00

MoritzLaurer/deberta-v3-base-zeroshot-v2 (the local_encoder default) started returning 401 on an unauthenticated GET of its own model page -- gated or moved sometime after this project picked it. Discovered live 2026-09-06 when it took production down entirely (crash-loop, exit code 1).

The startup check (ensure_available) did exactly its job -- refused to boot rather than fail opaquely on the first request -- but the service still couldn't come up until the default itself was fixed.

Switched to facebook/bart-large-mnli: HuggingFace's own reference model for the zero-shot-classification pipeline, confirmed publicly accessible live (200, vs the old model's 401, checked with a browser user-agent to rule out bot-blocking). Larger than the old default (~407M vs ~184M params), but that's the honest tradeoff for "still works."

Also updated the confidence_threshold documentation in docs/local-models.md to note the [0.0, 1.0] probability requirement from #47 (a value like 80 used to be silently accepted and broke classification the same way).

Full suite: 1503 passed, 0 failures.

`MoritzLaurer/deberta-v3-base-zeroshot-v2` (the `local_encoder` default) started returning 401 on an unauthenticated GET of its own model page -- gated or moved sometime after this project picked it. Discovered live 2026-09-06 when it took production down entirely (crash-loop, exit code 1). The startup check (`ensure_available`) did exactly its job -- refused to boot rather than fail opaquely on the first request -- but the service still couldn't come up until the default itself was fixed. Switched to `facebook/bart-large-mnli`: HuggingFace's own reference model for the `zero-shot-classification` pipeline, confirmed publicly accessible live (200, vs the old model's 401, checked with a browser user-agent to rule out bot-blocking). Larger than the old default (~407M vs ~184M params), but that's the honest tradeoff for "still works." Also updated the confidence_threshold documentation in docs/local-models.md to note the `[0.0, 1.0]` probability requirement from #47 (a value like `80` used to be silently accepted and broke classification the same way). Full suite: 1503 passed, 0 failures.

alee added 1 commit 2026-09-07 00:34:21 +00:00

fix(classifier): local_encoder default model is gated, switch to bart-large-mnli 1f74895910

MoritzLaurer/deberta-v3-base-zeroshot-v2 (the local_encoder default)
started returning 401 on an unauthenticated GET of its own model page
-- gated or moved sometime after this project picked it, discovered
live 2026-09-06 when it took production down entirely (crash-loop,
exit code 1). The startup check (ensure_available) did its job --
refused to boot rather than fail opaquely on the first request -- but
the service still couldn't come up until the default was fixed.

Switched to facebook/bart-large-mnli: HuggingFace's own reference
model for the zero-shot-classification pipeline, confirmed publicly
accessible live. Larger than the old default (~407M vs ~184M params)
but that's the honest tradeoff for "still works."

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VRQXz5SYZYVWscxS1QqF6U

alee merged commit 5b75fa3f81 into main

2026-09-07 00:34:48 +00:00

alee referenced this issue from a commit

2026-09-07 00:34:50 +00:00

Merge pull request 'fix(classifier): local_encoder default model is gated, switch to bart-large-mnli' (#48) from fix/local-encoder-default-model into main

alee referenced this issue from a commit

2026-09-07 00:48:37 +00:00

fix(deploy): redirect HF_HOME into the repo, same crash on a fresh install

alee referenced this issue from a commit

2026-09-07 01:28:27 +00:00

fix(deploy): redirect HF_HOME into the repo, same crash on a fresh install

alee referenced this pull request