Today is a big day for us. We are shipping two models at once: Supra-50M Base and Supra-50M Instruct. This is the first release under our Scaling Up Plan, and the jump from 8M to 50M parameters is the largest single leap we have made so far. The Base model was pretrained from scratch on 20 billion tokens. The Instruct model was then fine-tuned on top of it for chat and instruction following. Two very different use cases, one release.

Supra-50M Base BASE

The Base model is a raw pretraining run. No instructions, no chat format, just a Llama-architecture decoder trained to predict the next token on high-quality educational text. The full config:

// supra-50m base — model config Parameters         → ~50M (51.8M)
Architecture      → Llama (decoder-only)
Vocab size        → 32,000 (custom BPE)
Hidden size       → 512
Intermediate size  → 1,408
Layers            → 12
Attention heads    → 8 (GQA: 4 KV heads)
Context length     → 1,024 tokens
Training tokens    → 20,000,000,000 (20B)
Dataset           → Fineweb-Edu (sample-100BT)
Final loss        → 3.259
Precision         → bfloat16

The tokenizer was trained from scratch on 500,000 documents from Fineweb-Edu using Byte-Level BPE with a 32,000 vocabulary. Training ran on a single GPU with torch.compile enabled, cosine LR schedule, and AdamW fused optimizer.

Benchmarks

We benchmarked the Base model against GPT-2 (124M), SmolLM-135M, and OpenELM-270M. Supra-50M uses 2 to 5 times fewer parameters than every model in the comparison.

Benchmark Supra-50M GPT-2 (124M) SmolLM (135M) OpenELM (270M)
Parameters 50M 124M (2.5×) 135M (2.7×) 270M (5.4×)
BLiMP (linguistics) 76.3% 63.0% 69.8% N/A
SciQ (science) 77.2% 53.2% 73.4% 84.7%
ARC-Easy (knowledge) 52.2% 42.0% 49.2% 45.1%
PIQA (logic) 62.2% 63.0% 67.3% 69.8%
HellaSwag (context) 31.8% 29.5% 42.0% 46.7%

BLiMP at 76.3% beats GPT-2 by over 13 points and SmolLM by 6.5 points, with a model that is 2.5 to 2.7 times smaller. ARC-Easy at 52.2% beats all three comparison models including those much larger. The weaker numbers on HellaSwag and PIQA reflect the longer-context reasoning those tasks require, which is an area to push on next.

Supra-50M Instruct INSTRUCT

The Instruct model takes the Base weights and fine-tunes them for instruction following using the full Alpaca-Cleaned dataset for 4 epochs. The prompt format used at training time, which you need to match at inference time, looks like this:

// instruct prompt template Below is an instruction that describes a task. Write a response
that appropriately completes the request.

### Instruction:
{your instruction here}

### Response:
{model output starts here}

If you send a raw prompt to the Instruct model without this format, you will get Base-model behavior. The template matters.

Base model outputs

// base model · prompt
"The main concept of physics is "
// output
The main concept of physics is iffy, and the idea that we can make things behave in a certain way. The most important part of physics is called quantum mechanics which states that all particles are made up of energy (energy) and matter (matter). In physics, there are two types of particles: elementary particles and exotic ones...
// base model · prompt
"Once upon a time, "
// output
Once upon a time, I was so excited about the new school year and wanted to make some changes in my life. I had been looking for ways to help me become more self-aware. As an adult, I have always felt that there is no one way of doing things without thinking first...

Instruct model output

// instruct model · user
What is AI?
// response
AI refers to the simulation of human intelligence and creativity in machines, which are designed to mimic or surpass human cognition and reasoning. It is the process by which AI systems can solve problems, make predictions, or automate processes that typically require human-level thinking and problem solving. AI technology has been rapidly growing and evolving over time, with many potential benefits for society as a whole...

The difference between Base and Instruct outputs is clear. The Base model completes text naturally. The Instruct model answers questions with structure, numbered lists, and a coherent response format. That is exactly what SFT is supposed to do.

How to run

BASE Standard pipeline, same pattern as previous Supra Mini releases:

from transformers import pipeline
import torch

print("[*] Loading Supra-50M Base...")
pipe = pipeline(
    "text-generation",
    model="SupraLabs/Supra-50M-Base",
    device_map="auto",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)

def generate_text(prompt, max_new_tokens=150):
    result = pipe(
        prompt,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        temperature=0.5,
        top_k=25,
        top_p=0.9,
        repetition_penalty=1.2,
        pad_token_id=pipe.tokenizer.pad_token_id,
        eos_token_id=pipe.tokenizer.eos_token_id
    )
    return result[0]['generated_text']

print(generate_text("The importance of education is"))

INSTRUCT The Instruct model needs the prompt template and a slightly different setup:

import os, warnings
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
warnings.filterwarnings("ignore", category=UserWarning, module="transformers")

import torch
from transformers import pipeline, AutoTokenizer, logging
logging.set_verbosity_error()

MODEL_ID = "SupraLabs/Supra-50M-Instruct"
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, clean_up_tokenization_spaces=False)
pipe = pipeline(
    "text-generation",
    model=MODEL_ID,
    tokenizer=tokenizer,
    device_map="auto",
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32
)

def build_prompt(instruction, input_text=""):
    if input_text.strip():
        return ("Below is an instruction that describes a task, paired with an input "
            "that provides further context. Write a response that appropriately "
            "completes the request.\n\n"
            f"### Instruction:\n{instruction}\n\n"
            f"### Input:\n{input_text}\n\n### Response:\n")
    return ("Below is an instruction that describes a task. Write a response that "
        "appropriately completes the request.\n\n"
        f"### Instruction:\n{instruction}\n\n### Response:\n")

def generate(instruction, input_text=""):
    result = pipe(build_prompt(instruction, input_text),
        max_new_tokens=512, do_sample=True, temperature=0.7,
        top_k=50, top_p=0.9, repetition_penalty=1.15,
        pad_token_id=pipe.tokenizer.pad_token_id,
        eos_token_id=pipe.tokenizer.eos_token_id,
        return_full_text=False)
    return result[0]['generated_text'].strip()

print(generate("Explain what a neural network is."))

What comes next

Supra-50M is the first step of the Scaling Up Plan. The Base model showed that our training setup and custom tokenizer scale well, and the Instruct fine-tune showed that the Base weights are a solid foundation for supervised learning. The next goals are pushing past 50M, experimenting with longer context, and improving HellaSwag and PIQA numbers where we are still behind models with more parameters.

Both models are live on HuggingFace right now. Go build something.

// links Base     → huggingface.co/SupraLabs/Supra-50M-Base
Instruct → huggingface.co/SupraLabs/Supra-50M-Instruct
License  → Apache 2.0
Series   → Project Chimera / Supra collection on HuggingFace
#release #supra-50m #chimera #base-model #instruct #sft #tinyml #llama #open-source #fineweb-edu