Raven Graves ravengraves

Local LLM stem/numerical software challenge

Results from a subset of Terminal Bench 2.1 with problems selected to form a representative benchmark for engineering software problems.

Terminal Bench 2.1 Scores (Engineering Subset)

Model	Mean (%)	Pass@2 (%)	Runtime
Qwen3.8-Flash-Next-UD-IQ4_XS.gguf (f16, thinking medium, 30tk/s gen, 200tk/s pp)	86.7	90.0	11h 4m 28s
Qwen3.8-27B-UD-IQ4_XS.gguf f16 kv (2x 5060 ti 16gb)	86.7	96.7	7h 4m 53s

installing Parabola GNU/Linux-libre with encryption

Partitioning

First, find your disk

fdisk -l

Now, modify your disk. Set it as MBR, set a boot partition.

Install Vicuna (Windows)

This guide assumes you have a pretty beefy system. You probably want a Nvidia 3090 or 4090 GPU and 32 GB of RAM. You have to spend money to make money, as they say.

Install and run Transmission QT client
[CTRL] + [U] and paste the link: magnet:?xt=urn:btih:ZXXDAUWYLRUXXBHUYEMS6Q5CE5WA3LVA&dn=LLaMA
1. Select a folder to download into
To only install LLaMA 7b and 13b:

	FROM ubuntu:24.04
	ENV DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC
	RUN apt-get update && apt-get install -y --no-install-recommends \
	git ca-certificates curl sudo xvfb mesa-utils \
	libgl1-mesa-dri libxcb-cursor0 libxkbcommon-x11-0 \
	libxcb-icccm4 libxcb-keysyms1 libxcb-shape0 \
	libxcb-render-util0 libxcb-xinerama0 qt6-wayland \
	&& rm -rf /var/lib/apt/lists/*

	# FreeCAD's own CI dependency installer (adds Neon repo for matched Qt6/PySide6)

	# nix develop .#cuda
	# nix develop .#intel-cuda
	# nix devolop .#rocm
	# nix develop .#cpu
	{
	description = "llama.cpp manual build environments (CUDA / ROCm + clang)";

	inputs = {
	nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable";
	flake-utils.url = "github:numtide/flake-utils";

	#include <stdio.h>
	#include <stdlib.h>
	#include <stdint.h>
	#include <stddef.h>
	#include <math.h>
	#include <time.h>
	#include <string.h>

	#if defined(__GNUC__) \|\| defined(__clang__)
	# define RESTRICT __restrict__

	#include <stdio.h>
	#include <stdlib.h>
	#include <stdint.h>
	#include <string.h>
	#include <math.h>
	#include <time.h>

	#if defined(__GNUC__) \|\| defined(__clang__)
	# define RESTRICT __restrict__
	# define INLINE static inline __attribute__((always_inline))

	# convert an enpass .csv to keepassxc input .csv format
	# Steps:
	# 1. Export your enpass to a csv: File > export > .csv > "export.csv" > Export
	# 2. Run this script from the same folder as export.csv
	# 3. Import "export_migrated.csv" into keepassxc.
	# 4. Delete "export.csv" and "export_migrated.csv" files
	import csv
	with open("export.csv", "r") as f:
	data = list(csv.reader(f))

	desc:Variable-TET Note Mapper (MPE)

	// Input MIDI note = TET step index.
	// Input note 0 maps to the selected base MIDI note.
	// Uses MPE member channels 2-16.
	// Supports only up to 16 notes at a time.

	// match pitch bend range with synth; usually 2.0
	slider1:2<1,12,1>Pitch Bend Range (semitones).
	slider2:16<1,72,1>TET Divisions per Octave