README.md

---
title: MiPLM Inference
emoji: 🧬
colorFrom: indigo
colorTo: blue
sdk: docker
pinned: false
app_port: 7860
---

# MiPLM inference API

FastAPI service that serves the MiPLM ESM-2 8M variants as a small REST API.

## Endpoints

| Method | Path        | Body                              | Returns                                  |
|--------|-------------|-----------------------------------|------------------------------------------|
| GET    | `/`         | —                                 | model list, device, max length           |
| GET    | `/health`   | —                                 | liveness + which models are loaded       |
| POST   | `/predict`  | `{sequence, model}`               | per-position 20-AA softmax `[L, 20]`     |
| POST   | `/embed`    | `{sequence, model}`               | mean-pooled last-layer embedding         |
| POST   | `/mutation` | `{sequence, model}`               | wildtype-marginal Δlog-likelihood `[L, 20]` |

`model` is one of `miplm-ce` (the cross-entropy 8M variant we trained, surfaced in the
demo as the "ESM-2" baseline) or `miplm-blosum` (BLOSUM soft-label 8M variant, surfaced
as "MiPLM"). The MSA variant is not trained yet. Sequences must be `[A-Za-z]+` and
≤ 1024 residues.

## Example

```bash
curl -X POST $SPACE_URL/predict \
  -H "Content-Type: application/json" \
  -d '{"sequence": "MKTAYIAKQRQISFVKSHFSRQLEERLGLIE", "model": "miplm-blosum"}'
```