---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
  - babelbit
  - bittensor
  - utterance-prediction
---

# Babelbit Miner Model

A fine-tuned utterance prediction model for the Babelbit subnet (netuid 59) on the Bittensor network.

---

## What this model does

Given a partial utterance prefix and optional conversation context, the model predicts the most natural and complete continuation of the utterance.

The subnet evaluates predictions on:
- **Lexical similarity** — exact word overlap with the ground truth
- **Semantic similarity** — meaning-level match
- **Earliness** — how early in the utterance the prediction was correct

---

## Usage

### Direct inference

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.float16)

system = (
    "You are a helpful assistant that completes the current utterance naturally and succinctly. "
    "Return only the completed utterance text without quotes or extra commentary."
)
prefix = "Hi - how"
context = ""

if context:
    user = f"Context:\n{context}\n\nContinue the utterance that begins with:\n{prefix}"
else:
    user = f"Continue the utterance that begins with:\n{prefix}"

messages = [
    {"role": "system", "content": system},
    {"role": "user", "content": user},
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=64, do_sample=False)

decoded = tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(decoded)
# → "Hi - how are you?"
```

### Validator API request format

Validators query the miner at `POST /predict`:

```json
{
  "index": "<session-uuid>",
  "step": 3,
  "prefix": "Hi - how",
  "context": "",
  "done": false
}
```

Expected response:

```json
{
  "prediction": "Hi - how are you?"
}
```

---

## Limitations

- Designed specifically for the Babelbit subnet utterance prediction task.
- Best results on conversational English; performance varies on other languages.
- Predictions are evaluated on partial prefixes — shorter prefixes are inherently harder.