SEN-AGI/sable-2-90m-preview

Sable 2 Preview - 90M

Sable 2 Preview is an early test, mid-trained version of the Sable 2 series. This checkpoint is released to test training stability and tokenizer performance and is not the final model.

Model Details

• Model ID: SEN-AGI/sable-2-90m-preview
• Developer: SEN-AGI
• Release Date: August 6, 2026 - Thursday
• Model Type: Causal Language Model
• Architecture: LLaMA
• Parameters: 89.19M
• Vocabulary Size: 24k
• Tokenizer: Custom BPE
• Status: Early Test / Mid-Trained
• Training Hardware: Google Colab T4 GPU
• Training Cost: $0

Intended Use

For research, experimentation, and educational use. This is a preview model and not suitable for production use. It may produce incoherent, repetitive, or factually incorrect outputs.

How to Use

Installation

pip install transformers torch

Text Generation with Pipeline

from transformers import pipeline

generator = pipeline(
    "text-generation",
    model="SEN-AGI/sable-2-90m-preview",
    tokenizer="SEN-AGI/sable-2-90m-preview"
)

output = generator(
    "The future of small language models is",
    max_new_tokens=64,
    temperature=0.7,
    top_k=50,
    do_sample=True,
    repetition_penalty=1.2
)

print(output[0]["generated_text"])

Manual Generation with Auto Classes

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("SEN-AGI/sable-2-90m-preview")
model = AutoModelForCausalLM.from_pretrained("SEN-AGI/sable-2-90m-preview")

prompt = "In 2026, researchers discovered that"
inputs = tokenizer(prompt, return_tensors="pt")

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=64,
        temperature=0.7,
        top_k=50,
        do_sample=True,
        repetition_penalty=1.2,
        pad_token_id=tokenizer.eos_token_id
    )

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

⚠️ Note: As a mid-trained 90M preview model, outputs will frequently be incoherent or repetitive. Use higher repetition_penalty (1.2–1.5) and moderate temperature (0.6–0.8) for best results during evaluation.

Limitations

• Mid-trained checkpoint with limited capabilities
• Small 89.19M parameter count limits reasoning and knowledge
• 24k vocab may cause fragmentation on rare words
• Not instruction-tuned and not aligned with RLHF
• May hallucinate or show bias