Raptor is a 1.027B-parameter decoder-only causal language model with approximately 404M active parameters per token. It uses multi-head latent attention and six SwiGLU experts per layer with top-2 routing.

This revision contains the English instruction-tuned checkpoint. It was initialized from the Raptor step-35,000 pretrained checkpoint and supervised fine-tuned for one epoch on a curated SmolTalk mixture. The retained checkpoint is SFT step 7,500, selected by validation loss.

Architecture

18 layers
hidden size 1,024
latent attention dimension 256
16 attention heads
six experts per layer, top-2 routing
expert hidden size 2,816
context length 2,048
35,000-token SentencePiece vocabulary
1.027B total parameters, about 404M active per token

Fine-tuning

Base checkpoint: pretraining step 35,000
Training examples: 511,721
Validation examples: 2,000
SFT epochs: 1
Best checkpoint: step 7,500
Best validation loss: 1.0597
Best validation perplexity: 2.8855
Training format: assistant-only loss over ### User: and ### Assistant: conversations

Usage

The architecture and tokenizer use custom code, so loading requires trust_remote_code=True.

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Voyager466920/Raptor"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{"role": "user", "content": "What is the capital of France?"}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)
output = model.generate(
    inputs,
    max_new_tokens=128,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
)
print(tokenizer.decode(output[0, inputs.shape[1]:], skip_special_tokens=True))

Limitations

This is an experimental 1B-scale model and may fail simple reasoning or arithmetic tasks.
Multi-turn memory and role consistency are unreliable.
Responses may become verbose, repetitive, inaccurate, biased, or unsafe.
The model is English-focused. The tokenizer has poor Korean coverage and maps many Korean words to the unknown token.
The architecture currently recomputes the full prefix during generation and does not implement a KV cache.

License

No model license has been selected yet. Public availability does not grant additional usage rights beyond applicable law.