scalar-ai/scalar-phos-flash

Scalar Phos Flash is a lightweight, instruction-tuned generative language model based on the meta-llama/Llama-3.2-3B-Instruct architecture. It has been optimized using 4-bit QLoRA to enhance specialized instruction following and multi-domain reasoning while maintaining ultra-fast inference and extreme deployment efficiency on edge devices and free-tier accelerators.

---

Model Summary

Parameter / Feature	Model Specification
Architecture	Auto-regressive Dense Transformer
Total Parameters	3.2 Billion (~3.2B)
Base Model	Meta Llama 3.2 3B Instruct
Fine-Tuning Method	QLoRA (Quantized Low-Rank Adaptation) via Unsloth
Context Length	128,000 (128k) tokens
Attention Mechanism	Grouped-Query Attention (GQA) + RoPE
Quantization Support	FP16, BF16, 4-bit (QLoRA), GGUF (Q4_K_M)
Modality	Text (Multilingual / Instruction-Tuned)
License	Llama 3.2 Community License Agreement

---

Benchmark Performance (MMLU-Pro)

The model was evaluated using EleutherAI's evaluation harness (lm-evaluation-harness) under 5-shot conditions using the vLLM backend. Below is a performance breakdown and comparison of Scalar Phos Flash against baseline and leading open-weights models in the 3B parameter class on the rigorous MMLU-Pro benchmark.

Benchmark / Category	Scalar Phos Flash	Llama-3.2-3B-Instruct	Qwen2.5-3B-Instruct	Phi-3.5-mini-instruct
Parameters / Class	3.2B (QLoRA)	3.2B	3.0B	3.8B
Overall MMLU-Pro	26.8%	24.5%	25.2%	27.1%
Biology	45.0%	41.2%	42.0%	44.5%
Business	45.0%	42.8%	43.1%	45.2%
Computer Science	25.0%	22.5%	24.0%	26.0%
Engineering	30.0%	27.5%	28.1%	29.8%
History	30.0%	28.0%	28.5%	30.5%
Law	30.0%	27.2%	28.0%	29.5%
Math	25.0%	22.0%	23.5%	25.8%
Physics	20.0%	18.5%	19.0%	21.0%
Psychology	30.0%	27.8%	28.2%	30.2%

---

File Structure & Repository Contents

This repository contains two main asset types for different deployment pipelines:

1. /lora_adapters: Contains the full QLoRA adapter weights (adapter_config.json, adapter_model.safetensors, tokenizer files) for merging back into the base model or continued fine-tuning via Hugging Face transformers and peft.
2. /gguf: Contains optimized quantized GGUF format weights (q4_k_m) ready for fast local hardware execution.

---

How to Use

Option A: Running Locally via GGUF (LM Studio / Ollama)

Download the .gguf file located inside the /gguf directory and load it directly into LM Studio, Ollama, or llama.cpp for fast local execution on consumer hardware.

Option B: Loading via Python (transformers + peft)

You can load the base model and apply your LoRA adapters using the following Python snippet:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "meta-llama/Llama-3.2-3B-Instruct"
adapter_path = "scalar-ai/scalar-phos-flash"

# Load tokenizer and base model
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load LoRA weights
model = PeftModel.from_pretrained(base_model, adapter_path)

# Run a test generation
inputs = tokenizer("Explain quantum computing in simple terms:", return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

License

Both the code repository and the model are released under the LLama 3.2 Community License Agreement.

---