Skip to content

feat(model-ports): add modernbert_sentiment - #211

Open
DarthCeltic wants to merge 2 commits into
aifoundry-org:mainfrom
DarthCeltic:port-modernbert-sentiment
Open

feat(model-ports): add modernbert_sentiment#211
DarthCeltic wants to merge 2 commits into
aifoundry-org:mainfrom
DarthCeltic:port-modernbert-sentiment

Conversation

@DarthCeltic

Copy link
Copy Markdown
Contributor

Adds modernbert_sentiment classification port to llama.cpp-et, with a model-ports claim. Artifact was already converted and hosted earlier this session (GitHub Release) but never had a claim filed until this audit caught it.

Recovers an already-converted, already-hosted artifact from earlier
this session (GitHub Release model-port-modernbert-sentiment-v1) that
never had a model-ports claim filed -- caught during a full re-audit
of all open PRs and releases. arch=modern-bert, distinct from bert
(distilbert_sst2, already claimed) and from every other bert-family
variant (nomic-bert, jina-bert-v2/v3, neo-bert, eurobert -- all still
unclaimed).

Verified live: built llama-server from a plain CPU-only (GGML_ET=OFF)
configuration of the vendored source, loaded the GGUF, and got a real
response from POST /rerank (3-way sentiment classifier head loads
correctly: negative/neutral/positive). Same verification tier as
distilbert_sst2/roberta_sst2 earlier in this campaign.

Registered only in artifacts.json, not benchmark_config.json -- same
reason as distilbert_sst2: the shared runner only speaks
/completion and /v1/chat/completions, not /rerank, so an automated
"board": true entry would be misleading without a maintainer adding
/rerank support to that protected script.
@DarthCeltic
DarthCeltic requested a review from AFOliveira as a code owner July 25, 2026 17:42
@github-actions github-actions Bot added track: model-ports Global approved model ports prize track: community Community support and ecosystem contributions labels Jul 25, 2026
@AFOliveira

Copy link
Copy Markdown
Member

ModernBERT currently has an artifact, recipe and claim but no registered board benchmark or committed deterministic sentiment oracle. Please add the benchmark, pinned test inputs + expected labels/logits and threshold, and a successful ET-SoC1 result. Maintainers will then review/adopt the identity and protected validation contract.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

track: community Community support and ecosystem contributions track: model-ports Global approved model ports prize

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants