README.md

---
license: mit
pipeline_tag: text2text-generation
tags:
- text2text-generation
- paraphrase
- t5
base_model: humarin/chatgpt_paraphraser_on_T5_base
---

# humarin-paraphraser-handler

Custom HF Inference Endpoint handler that wraps
[`humarin/chatgpt_paraphraser_on_T5_base`](https://huggingface.co/humarin/chatgpt_paraphraser_on_T5_base)
with explicit `T5ForConditionalGeneration` + diverse beam search,
so HF auto-routing cannot mis-load it as a causal LM.

## Why this exists

When deployed directly, HF Inference Endpoints routed the base model to
vLLM / TGI, which treats T5 as causal generation and produces broken output
(echoes the input prompt instead of returning a paraphrase).

This repo only ships `handler.py` + `requirements.txt`; the model weights
are loaded at runtime from the base repo.

## API

POST with JSON body:

```json
{
  "inputs": "The product is excellent and works well.",
  "parameters": {
    "num_beams": 5,
    "num_beam_groups": 5,
    "num_return_sequences": 5,
    "diversity_penalty": 3.0,
    "repetition_penalty": 10.0,
    "no_repeat_ngram_size": 2,
    "max_length": 128
  }
}
```

Returns 5 paraphrase candidates as `[{"generated_text": ...}, ...]`.