Modelpublic

SupraLabs/Supra-1.5-50M-instruct-exp-gguf

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

18likes792downloads

Model Card

<h1 align="center">Supra-1.5 Instruct • Experimental Chat Tune — GGUF</h1>

Supra-1.5 Instruct

GGUF quantizations of SupraLabs/Supra-1.5-50M-instruct-exp, an experimental 50M-parameter instruction-tuned model by SupraLabs, part of Project Chimera.

Run it entirely on CPU, low-VRAM GPUs, or embedded hardware. No cloud required.

Note: This is an experimental model. Do not use in production.

📦 Available Quantizations

Bits	Quantization	Size
1-bit	`Q1_0`	19.6 MB
1-bit	`TQ1_0`	25.1 MB
2-bit	`Q2_K`	28.8 MB
2-bit	`TQ2_0`	26.4 MB
3-bit	`IQ3_S`	31 MB
3-bit	`Q3_K_S`	31 MB
3-bit	`IQ3_M`	31.7 MB
3-bit	`Q3_K_M`	32.7 MB
3-bit	`Q3_K_L`	33.8 MB
4-bit	`IQ4_XS`	33.8 MB
4-bit	`Q4_K_S`	35.7 MB
4-bit	`IQ4_NL`	34.7 MB
4-bit	`Q4_0`	34.5 MB
4-bit	`Q4_1`	36.8 MB
4-bit	`Q4_K_M`	37.4 MB
5-bit	`Q5_K_S`	39.5 MB
5-bit	`Q5_0`	39 MB
5-bit	`Q5_1`	41.2 MB
5-bit	`Q5_K_M`	41 MB
6-bit	`Q6_K`	45.8 MB
8-bit	`Q8_0`	56.2 MB
16-bit	`BF16`	105 MB
16-bit	`F16`	105 MB
32-bit	`F32`	208 MB

`Q4_K_M` — Usable, not recommended unless device is compute-constrained. `Q8_0` — Perfect size/performance!. `Q2_K` — ultra-constrained devices (not reccomended!).

🚀 Quick Start

llama.cpp

bash

# Download
huggingface-cli download SupraLabs/Supra-1.5-50M-instruct-exp-gguf \
  --include "*.Q4_K_M.gguf" \
  --local-dir ./

# Run
./llama-cli \
  -m supra-1.5-50m-instruct-exp-Q4_K_M.gguf \
  -p "### Instruction:\nWhat is machine learning?\n\n### Response:\n" \
  -n 256 \
  --temp 0.7 \
  --repeat-penalty 1.15

Ollama

bash

ollama run hf.co/SupraLabs/Supra-1.5-50M-instruct-exp-gguf:Q4_K_M

Python (llama-cpp-python)

python

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="SupraLabs/Supra-1.5-50M-instruct-exp-gguf",
    filename="*Q4_K_M.gguf",
    n_ctx=1024,
    verbose=False,
)

def chat(instruction: str, input_text: str = "") -> str:
    if input_text.strip():
        prompt = (
            "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"
        )
    else:
        prompt = (
            "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"
        )
    output = llm(prompt, max_tokens=256, temperature=0.7, top_k=50, top_p=0.9, repeat_penalty=1.15)
    return output["choices"][0]["text"].strip()

print(chat("Explain what artificial intelligence is."))

💬 Prompt Format

This model uses the Alpaca Chat Format:

Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Response:

With optional input:

Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{instruction}

### Input:
{input}

### Response:

🏆 Benchmarks

Supra-1.5-50M-instruct-exp achieves superior performance within the 50M-parameter class, with a consistent BLiMP score of 67.4.

Key findings from evaluation:

—Scientific/factual tasks perform best under raw inference (no normalization)
—Math and logical reasoning benefit from normalized inference
—Top syntactic categories: structural dependency tracking, complex clausal configurations, and subtle syntactic error detection — performing at near-flawless precision
—Hardest categories: advanced binding phenomena and morphological agreement edge cases, reflecting known limits of 50M-class architectures

For full benchmark charts and BLiMP probe analysis, see the base model card.

🧠 Model Architecture

Property	Value
Architecture	Llama (decoder-only)
Parameters	~50M
Vocabulary	32,000 (custom BPE)
Context length	5,120 tokens
Hidden size	512
Layers	12
Attention heads	8 (GQA: 4 KV heads)
Base model	SupraLabs/Supra-1.5-50M-Base-exp
License	Apache 2.0

🔗 Related Models

Model	Description
Supra-1.5-50M-Base-exp	Pretrained base (v1.5)
Supra-1.5-50M-instruct-exp	fp weights
Supra-50M-Base	v1.0 pretrained base
Supra-50M-Instruct	v1.0 instruct model
Supra-50M-Reasoning	Chain-of-thought reasoning variant

📄 License

Released under the Apache 2.0 License.

Credit goes to @QyrouNnet-AI