94 lines
2.5 KiB
Markdown
94 lines
2.5 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
base_model: HuggingFaceTB/SmolLM2-360M-Instruct
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
tags:
|
||
|
|
- text-generation
|
||
|
|
- conversational
|
||
|
|
- smollm2
|
||
|
|
- wvy
|
||
|
|
- safetensors
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
library_name: transformers
|
||
|
|
---
|
||
|
|
|
||
|
|
# BbyWVY-360m
|
||
|
|
|
||
|
|
BbyWVY-360m is an experimental conversational language model based on
|
||
|
|
`HuggingFaceTB/SmolLM2-360M-Instruct`.
|
||
|
|
|
||
|
|
This checkpoint was first instruction-tuned for the WVY chat identity and then
|
||
|
|
continued-pretrained on a user-provided, user-message-only corpus. Assistant
|
||
|
|
messages and raw training data are not included in this repository.
|
||
|
|
|
||
|
|
## Intended Behavior
|
||
|
|
|
||
|
|
WVY is tuned to be conversational, curious, and uncertainty-aware:
|
||
|
|
|
||
|
|
- admit when it does not know something deeply
|
||
|
|
- ask useful follow-up questions
|
||
|
|
- build conclusions from what the user explains
|
||
|
|
- ask for verification instead of acting like it knows everything
|
||
|
|
|
||
|
|
## Local Usage
|
||
|
|
|
||
|
|
```python
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
|
||
|
|
model_id = "StarpowerTechnology/BbyWVY-360m"
|
||
|
|
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||
|
|
model = AutoModelForCausalLM.from_pretrained(model_id)
|
||
|
|
|
||
|
|
messages = [
|
||
|
|
{"role": "system", "content": "u are WVY. be curious, honest, and conversational."},
|
||
|
|
{"role": "user", "content": "what do u know about quantum physics?"},
|
||
|
|
]
|
||
|
|
|
||
|
|
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
||
|
|
inputs = tokenizer(prompt, return_tensors="pt")
|
||
|
|
outputs = model.generate(**inputs, max_new_tokens=120, temperature=0.7, top_p=0.95, do_sample=True)
|
||
|
|
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
|
||
|
|
```
|
||
|
|
|
||
|
|
## Hugging Face API
|
||
|
|
|
||
|
|
If this model is enabled through Hugging Face Inference Providers or a dedicated
|
||
|
|
Inference Endpoint, it can be called with the Hugging Face router:
|
||
|
|
|
||
|
|
```python
|
||
|
|
import os
|
||
|
|
from openai import OpenAI
|
||
|
|
|
||
|
|
client = OpenAI(
|
||
|
|
base_url="https://router.huggingface.co/v1",
|
||
|
|
api_key=os.environ["HF_TOKEN"],
|
||
|
|
)
|
||
|
|
|
||
|
|
completion = client.chat.completions.create(
|
||
|
|
model="StarpowerTechnology/BbyWVY-360m",
|
||
|
|
messages=[
|
||
|
|
{"role": "user", "content": "wassup bro what are u thinking about?"}
|
||
|
|
],
|
||
|
|
max_tokens=120,
|
||
|
|
temperature=0.7,
|
||
|
|
)
|
||
|
|
|
||
|
|
print(completion.choices[0].message.content)
|
||
|
|
```
|
||
|
|
|
||
|
|
## Training Note
|
||
|
|
|
||
|
|
Continued-pretraining settings:
|
||
|
|
|
||
|
|
- blocks: 9587
|
||
|
|
- block size: 2048 tokens
|
||
|
|
- epochs: 1
|
||
|
|
- learning rate: 2e-6
|
||
|
|
- trainable layers: last 4 transformer layers
|
||
|
|
- raw assistant messages excluded
|
||
|
|
|
||
|
|
This is an experimental checkpoint and may require additional SFT/alignment
|
||
|
|
passes to reduce repetition and tighten the curiosity loop.
|