elmorshedyahmed/Ahmed

Hugging Face Endpoint Template: mluke-store-judge

This folder contains a custom Inference Endpoint handler for running studio-ousia/mluke-large as a lightweight term-typing judge for Rosetta experiments.

Why this exists:

• the raw Hugging Face deployment for mLUKE exposes a fill-mask interface
• for Rosetta we need a typed classification interface for preserve experiments
• this wrapper converts candidate term + store context into scored label outputs

Files:

• handler.py: custom endpoint handler implementing EndpointHandler
• requirements.txt: endpoint dependencies

Expected request body

{
  "inputs": {
    "candidate": "Third Najm",
    "context": "Third Najm line: Syria-focused, spans hoodie/crewneck/tee/long-sleeve in black and white. Quoted poetic titles are used as named design concepts.",
    "labels": [
      "brand_name",
      "product_line",
      "named_work",
      "person_reference",
      "cultural_item",
      "nationality",
      "modifier"
    ]
  }
}

Expected response body

{
  "model_id": "studio-ousia/mluke-large",
  "candidate": "Third Najm",
  "top_label": "named_work",
  "results": [
    {
      "label": "named_work",
      "descriptor": "title",
      "raw_score": -2.13,
      "confidence": 0.44
    }
  ]
}

Deployment steps

1. Create a small Hugging Face repo under your account, for example:
   • new-account/mluke-store-judge
2. Upload handler.py and requirements.txt from this folder to that repo.
3. Create an Inference Endpoint from that repo using the default container type with task Custom.
4. Set endpoint env var:
   • MODEL_ID=studio-ousia/mluke-large
5. After deploy, save the endpoint URL locally as:
   • HF_MLUKE_ENDPOINT_URL=https://...your-endpoint-url...

Notes

• This is an experiment path only. It is not wired into the main Rosetta pipeline.
• The handler uses masked-LM pseudo-log-likelihood over label descriptors to score label fit.
• Default descriptor map:
  • brand_name -> brand
  • product_line -> collection
  • named_work -> title
  • person_reference -> person
  • cultural_item -> garment
  • nationality -> origin
  • modifier -> descriptor
• You can override that map by sending inputs.label_map in the request payload.