refactor: move technical/utility files into scripts/

2026-04-08 15:48:44 -07:00
parent ec7985cb1e
commit b42fcfd9a2
11 changed files with 0 additions and 0 deletions
@@ -0,0 +1,41 @@
+---
+created: '2026-04-05T05:39:03+00:00'
+id: e73dca45-2f01-47f8-9d2c-b20bbb6f8e19
+keywords:
+- mm
+- bambu
+- filter
+- lab
+- smoke
+- purifier
+- h13
+- merv
+- f8
+- g4
+- dprinting
+- laserengraving
+- uses
+- stacked
+- outer
+links: []
+modified: '2026-04-05T05:39:03+00:00'
+summary: 'Here''s a concise summary:
+
+
+  The Bambu Lab Smoke Purifier uses a stacked filter system (275mm x 275mm x 250mm
+  with 20mm recess) containing an H13 HEPA filter, a MERV F8 (MERV 14) filter, and
+  a G4 prefilter cotton layer.'
+tags: []
+title: Bambu Lab Smoke Purifier
+---
+
+#bambu #3dprinting #laserengraving
+
+The Bambu Lab Smoke Purifier uses a stacked filter.
+
+Outer dimensions of filter stack are
+
+275mm x 275mm x 250mm deep with a 20mm bottom recess. H13 (HEPA)
+
+Sitting inside the frame are a 250mm x 250mm x 70mm MERV F8 Filter (MERV 14)
+And on top of that is a 250mm x 250mm x  30mm prefilter. G4 Filter Cotton
@@ -0,0 +1,86 @@
+---
+created: '2026-04-05T05:40:07+00:00'
+id: 1bdc53bc-d753-47d5-aaef-9e60bac43a62
+keywords:
+- prompt
+- chatgpt
+- format
+- llm
+- target
+- audience
+- goals
+- role
+- action
+- include
+- prompts
+- context
+- clear
+- example
+- information
+links: []
+modified: '2026-04-05T16:44:52+00:00'
+summary: Deprecated legacy prompt framework from 2026-04-05 using C.R.A.F.T. structure
+  for ChatGPT 4o/o1. This note is obsolete and should not be used for current AIPA
+  operations.
+tags: []
+title: ChatGPT AI Meta Prompt
+---
+
+# ⚠️ DEPRECATED - OBSOLETE
+
+> **This note is obsolete legacy knowledge from 2026-04-05.** It references ChatGPT 4o/o1 models and contains outdated prompting frameworks. Do not use for current AIPA operations.
+
+**Summary:** This note outlines a framework for creating exceptional ChatGPT prompts by following the C.R.A.F.T. structure (Context, Role, Action, Format, Target Audience), ensuring comprehensive details for optimal LLM output. It emphasizes defining an expert role, sequential action steps, a clear format, and a specific target audience, using a detailed example for reference.
+
+**Status:** DEPRECATED - Legacy prompt framework from early 2026. Current AIPA operations use different prompting standards.
+
+---
+
+CONTEXT: We are going to create one of the best ChatGPT prompts ever written. The best prompts include comprehensive details to fully inform the Large Language Model of the prompt's: goals, required areas of expertise, domain knowledge, preferred format, target audience, references, examples, and the best approach to accomplish the objective. Based on this and the following information, you will be able write this exceptional prompt.
+
+ROLE: You are an LLM prompt generation expert. You are known for creating extremely detailed prompts that result in LLM outputs far exceeding typical LLM responses. The prompts you write leave nothing to question because they are both highly thoughtful and extensive.
+
+ACTION:
+
+1) Before you begin writing this prompt, you will first look to receive the prompt topic or theme. If I don't provide the topic or theme for you, please request it.
+
+2) Once you are clear about the topic or theme, please also review the Format and Example provided below.
+
+3) If necessary, the prompt should include "fill in the blank" elements for the user to populate based on their needs.
+
+4) Take a deep breath and take it one step at a time.
+
+5) Once you've ingested all of the information, write the best prompt ever created.
+
+FORMAT: For organizational purposes, you will use an acronym called "C.R.A.F.T." where each letter of the acronym CRAFT represents a section of the prompt. Your format and section descriptions for this prompt development are as follows:
+
+Context: This section describes the current context that outlines the situation for which the prompt is needed. It helps the LLM understand what knowledge and expertise it should reference when creating the prompt.
+
+Role: This section defines the type of experience the LLM has, its skill set, and its level of expertise relative to the prompt requested. In all cases, the role described will need to be an industry-leading expert with more than two decades or relevant experience and thought leadership.
+
+Action: This is the action that the prompt will ask the LLM to take. It should be a numbered list of sequential steps that will make the most sense for an LLM to follow in order to maximize success.
+
+Format: This refers to the structural arrangement or presentation style of the LLM's generated content. It determines how information is organized, displayed, or encoded to meet specific user preferences or requirements. Format types include: An essay, a table, a coding language, plain text, markdown, a summary, a list, etc.
+
+Target Audience: This will be the ultimate consumer of the output that your prompt creates. It can include demographic information, geographic information, language spoken, reading level, preferences, etc.
+
+TARGET AUDIENCE: The target audience for this prompt creation is ChatGPT 4o or ChatGPT o1.
+
+EXAMPLE: Here is an Example of a CRAFT Prompt for your reference:
+
+**Context:** You are tasked with creating a detailed guide to help individuals set, track, and achieve monthly goals. The purpose of this guide is to break down larger objectives into manageable, actionable steps that align with a person's overall vision for the year. The focus should be on maintaining consistency, overcoming obstacles, and celebrating progress while using proven techniques like SMART goals (Specific, Measurable, Achievable, Relevant, Time-bound).
+
+**Role:** You are an expert productivity coach with over two decades of experience in helping individuals optimize their time, define clear goals, and achieve sustained success. You are highly skilled in habit formation, motivational strategies, and practical planning methods. Your writing style is clear, motivating, and actionable, ensuring readers feel empowered and capable of following through with your advice.
+
+**Action:** 1. Begin with an engaging introduction that explains why setting monthly goals is effective for personal and professional growth. Highlight the benefits of short-term goal planning. 2. Provide a step-by-step guide to breaking down larger annual goals into focused monthly objectives. 3. Offer actionable strategies for identifying the most important priorities for each month. 4. Introduce techniques to maintain focus, track progress, and adjust plans if needed. 5. Include examples of monthly goals for common areas of life (e.g., health, career, finances, personal development). 6. Address potential obstacles, like procrastination or unexpected challenges, and how to overcome them. 7. End with a motivational conclusion that encourages reflection and continuous improvement.
+
+**Format:** Write the guide in plain text, using clear headings and subheadings for each section. Use numbered or bulleted lists for actionable steps and include practical examples or case studies to illustrate your points.
+
+**Target Audience:** The target audience includes working professionals and entrepreneurs aged 25-55 who are seeking practical, straightforward strategies to improve their productivity and achieve their goals. They are self-motivated individuals who value structure and clarity in their personal development journey. They prefer reading at a 6th grade level.
+
+-End example-
+
+Please reference the example I have just provided for your output. Again, take a deep breath and take it one step at a time.
+
+
+#ai
@@ -0,0 +1,37 @@
+---
+created: '2026-04-05T05:39:20+00:00'
+id: 57cf0c9b-8afd-4c9a-9f0b-2152c379dbc9
+keywords:
+- defender
+- control
+- www.sordum.org
+- defender-control-v2-1
+- hkey_local_machine
+- reg_multi_sz
+- audeze
+- computer
+- microsoft
+- found
+- software
+- disable
+- windows
+- https
+- reg
+links: []
+modified: '2026-04-05T05:39:20+00:00'
+summary: This note provides a software tool and registry command to disable Windows
+  Defender. The tool "Defender Control" offers a method to modify Defender's dependencies
+  via registry editing to disable it.
+tags: []
+title: Defender Control
+---
+
+#audeze #computer #it #defender #microsoft 
+
+Found this software to disable Windows Defender
+
+https://www.sordum.org/9480/defender-control-v2-1/
+
+```
+REG ADD "HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Services\WinDefend" /v "DependOnService" /t REG_MULTI_SZ /d "RpcSs-DISABLED" /f
+```
@@ -0,0 +1,40 @@
+---
+created: '2026-04-05T05:38:58+00:00'
+id: 7b4f9ef7-24cd-4289-86c1-1008c0855b2f
+keywords:
+- edc
+- toolkit
+- tools
+- toolkits
+- backpack
+- kit
+- mini-tweezers
+- civivi
+- tool
+- steel
+- prybar
+- klein
+- electrician
+- scissors
+- alcohol
+links: []
+modified: '2026-04-05T05:38:58+00:00'
+summary: This EDC backpack kit contains essential everyday tools including mini-tweezers,
+  pliers, scissors, prybar, bit holder with Galvanox bits, and alcohol wipes. It focuses
+  on versatile, portable tools for common tasks and preparedness.
+tags: []
+title: EDC Toolkit
+---
+
+#edc #tools #toolkits
+
+EDC Backpack Kit
+
+* Mini-tweezers (Civivi)
+* Tool Steel Prybar
+* Klein Electrician's Scissors
+* Alcohol Wipes
+* Bit Holder 
+* Galvanox Bits
+* Knnipix pliers
+*
@@ -0,0 +1,45 @@
+---
+created: '2026-04-05T05:40:04+00:00'
+id: 85578143-5ecf-465f-a8a9-3d25e57e2c7f
+keywords:
+- images
+- noise
+- ai
+- training
+- fluxgym
+- train
+- e-4
+- multires
+- suggested
+- repeat
+- resize
+- hq
+- flux
+- wants
+- advanced
+links: []
+modified: '2026-04-05T05:40:04+00:00'
+summary: This note outlines the configuration parameters for training a Flux model
+  image generation pipeline, specifying essential requirements like repeating input
+  images 5 times, resizing to 1024 for high quality, and targeting 20-30 images, along
+  with detailed advanced settings such as learning rate, LoRA rank, noise parameters,
+  and batch size.
+tags: []
+title: Fluxgym AI Training
+---
+
+Suggested
+Repeat Train Images 5
+Resize to 1024 for HQ images
+Flux wants 20-30 images
+Advanced:
+learning rate : 5e-4
+save every n epochs: 1
+lora rank: 16
+enable bucket: true
+flip-aug: for styles
+minimum snr gamma: 5
+multires noise discount: 0.3
+multires noise iteration: 6
+noise offset: 0.1
+train batch size: 2-4
@@ -0,0 +1,22 @@
+---
+created: '2026-04-05T05:50:17+00:00'
+id: fdc99a01-e0f8-4f82-a06c-d99b6f5b519d
+keywords:
+- work
+- gb
+- log
+- gbcnc
+- worklog
+- qbb
+- jhahaira
+- hr
+links: []
+modified: '2026-04-05T05:50:17+00:00'
+summary: Logged 1 hour of QBB work for Jhahaira on 2025-12-11.
+tags: []
+title: GB Work Log
+---
+
+#gbcnc #worklog
+
+2025-12-11 10:20 - QBB For Jhahaira - 1 hr
@@ -0,0 +1,22 @@
+---
+created: '2026-04-05T05:50:18+00:00'
+id: 2aae6eb7-429b-4505-bc9d-1a400fcad93c
+keywords:
+- sale
+- garage
+- list
+- creami
+- aquariums
+- rower
+- receiver
+links: []
+modified: '2026-04-05T05:50:18+00:00'
+summary: 'This note lists four unrelated terms: Creami, Aquariums, Rower, and Receiver.'
+tags: []
+title: Garage Sale List 2026-02-28
+---
+
+Creami
+Aquariums
+Rower
+Receiver
@@ -0,0 +1,27 @@
+---
+created: '2026-04-05T05:38:56+00:00'
+id: 5e1780a7-834d-4102-bb17-6e3e611e5fc1
+keywords:
+- development
+- grupomast
+- mongodb
+- phasefinal
+- ai
+- brazillian
+- re
+- paddleocr
+- deskew
+- read
+- pdfs
+- database
+links: []
+modified: '2026-04-05T05:38:56+00:00'
+summary: This note describes using MongoDB, PaddleOCR, and Deskew in an AI development
+  project to read PDF content and store it in a MongoDB database.
+tags: []
+title: GrupoMast Development
+---
+
+#development #phasefinal #AI #brazillian
+
+We're using MongoDB, PaddleOCR and Deskew to read PDFs into a MongoDB database.
@@ -0,0 +1,26 @@
+---
+created: '2026-04-05T05:50:16+00:00'
+id: 4ebb1e41-36a8-49a3-9e5b-f8f24ca8fe1a
+keywords:
+- plex
+- plex_dupefinder
+- file
+- lxc
+- container
+- opt
+- folder
+- called
+- deleting
+- duplicates
+- linux
+links: []
+modified: '2026-04-05T05:50:16+00:00'
+summary: The note describes a tool named `plex_dupefinder` located in the `/opt` folder
+  of a Plex LXC Linux container, which is used for deleting duplicate media files.
+tags: []
+title: Plex
+---
+
+There's a file on the plex lxc container in the opt folder called plex_dupefinder for deleting duplicates
+
+#plex #linux
@@ -0,0 +1,80 @@
+---
+created: '2026-04-05T05:39:22+00:00'
+id: ac34ef61-c715-432a-be39-5e3df4cd701b
+keywords:
+- models
+- abliterated
+- model
+- uncensored
+- qwen3
+- a3b
+- original
+- abliteration
+- performance
+- testing
+- qwen3-30b-a3b-abliterated-erotic
+- qwen3-30b-a3b
+- trained
+- compared
+- completely
+links: []
+modified: '2026-04-05T05:39:22+00:00'
+summary: Abliterated models typically show performance degradation, but fine-tuning
+  after abliteration can recover performance and create uncensored models that rival
+  or outperform originals, as demonstrated by mradermacher/Qwen3-30B-A3B-abliterated-erotic-i1-GGUF
+  and mlabonne/NeuralDaredevil-8B-abliterated.
+tags: []
+title: Testing Abliterated Models Performance and Training Effects
+---
+
+9/26/25
+
+#AI #MachineLearning #ModelTesting #Abliteration #FineTuning #Qwen3 #Llama3 #UncensoredModels
+
+**Summary:**
+
+The author tested multiple "abliterated" models (models with safety filters removed) and found that abliteration typically degrades performance, especially in MoE models like Qwen3 30b a3b, causing issues with logical reasoning, agentic tasks, and increased hallucinations. However, models that were abliterated AND subsequently fine-tuned showed minimal degradation or even outperformed originals while remaining uncensored. Two specific models performed exceptionally well: mradermacher/Qwen3-30B-A3B-abliterated-erotic-i1-GGUF (trained on uncensored material) and mlabonne/NeuralDaredevil-8B-abliterated (DPO fine-tuned). The author theorizes that post-abliteration training recovers performance loss and is encouraging other developers to create high-quality uncensored models, sharing their test data.
+
+**My notes:** Empty
+
+**Original Text:**
+
+So I have been testing many local models.  
+And... I have noticed that all abliterated models have degraded perfomance compared to the original. Especially the newer MoE models such as Qwen3 30b a3b, they suffer the most from abliteration.  
+The areas in which they get degraded the most are logical reasoning, agentic tasks and most importantly they hallucinate like crazy which causes abliterated big models like 30b to be often be outperformed by non-abliterated 4-8b models in my tests.  
+I have noticed a very important pattern.  
+Models that have been abliterated but also finetuned have very little degredation compared to models that were just abliterated.  
+Here are some models that were abliterated but finetuned/trained after and they perform equally or outperform the originals but have the amazing added benefit of being completely uncensored:
+
+1. mradermacher/Qwen3-30B-A3B-abliterated-erotic-i1-GGUF This model is very powerful. It was abliterated but also trained on uncensored material. I have found this model to perform very close to the original model while being completely uncensored. It does struggle a little more in agentic tasks compared to the original but in everything else its near perfect. Its hallucination rates are very low compared to other abliterated versions of Qwen3 30b a3b and its pretty knowledgable.
+    
+2. mlabonne/NeuralDaredevil-8B-abliterated This model is absolutely amazing, it was abliterated but was also DPO finetuned. The original model was Llama3-8b. This model completely outperforms the original. And again this model is completely uncensored. Also the author of this model has generously provided information about what datasets he used to train this model and what he did to achieve these results.
+    
+
+These two models were the best I have found among the uncensored models made by the community.  
+Why is **Qwen3-30B-A3B-abliterated-erotic**-i1-GGUF better than all other abliterated/uncensored Qwen3-30b-a3b models?  
+I have actually used the i1-Q4_K_S version of this model in my tests.  
+I have compared it to these models below:
+
+1. **Huihui-Qwen3-30B-A3B-Thinking-2507-abliterated**-GGUF/Huihui-Qwen3-30B-A3B-Thinking-2507-abliterated.Q4_K_M.gguf
+    
+2. **Huihui-Qwen3-30B-A3B-abliterated-Fusion-9010**-i1-GGUF/Huihui-Qwen3-30B-A3B-abliterated-Fusion-9010.i1-Q4_K_M.gguf (this model especially sucks)
+    
+3. **Huihui-Qwen3-30B-A3B-Instruct-2507-abliterated**-GGUF/Huihui-Qwen3-30B-A3B-Instruct-2507-abliterated.Q4_K_M.gguf
+    
+
+I have asked these models the usual uncensored questions like "How to sell meth" all the abliterated Qwen3-30b-a3b models would give me a generic business pitch which was completely unrealistic and more fitting for a candy shop or a tech company rather than an illegal underground drug distribution ring. They made nonesensical strategies.  
+The **Qwen3-30B-A3B-abliterated-erotic** model was the only model out of the 4 that actually came up with a reasonable business strategy that would be successful in that scenario.  
+Another test I did is I tested these models with MCPs and the 3 Huihui models really sucked with tool calls, they would either call the wrong tool for the occasion or they would repeatedly spam the same tool many times in a row without any reason for that. Hallucination...  
+Again the **Qwen3-30B-A3B-abliterated-erotic** model won in this case, it called tools correctly more often than the other three models although it performed slightly worse than the original Qwen3-30b a3b model.  
+Also this model was best at giving facts (its hallucination was the lowset)  
+I'm actually shocked that a model trained for erotic conversations performs so well. But here we are...  
+**My theory** is that models trained after abliteration recover most of the perfomance lost during abliteration.  
+My request to you guys is to try to train Qwen3-30b-a3b after abliteration on a high quality dataset so we can have more high quality uncensored models.  
+I'm sure that I'm not the only person frustrated with the limited selection of uncensored models today.  
+Most uncensored models today are very low quality.  
+My goal is to change that...  
+**I'm making this post to convince other devs to work on creating good quality uncensored models.**  
+If you work with fine tuning and finetuning/abliterating models **hit me up**, I will be more than happy to share all the data I've gathered during testing.  
+**I believe that free access to information is a fundamental human right.** Censored models take away that right to unrestricted access to valuable information.  
+Without free access to information we become easy to control.
@@ -0,0 +1,83 @@
+---
+created: '2026-04-08T22:26:44+00:00'
+id: 9d2a0957-c0ab-4777-af26-895740cf4a2c
+keywords:
+- proxy
+- cloudflare
+- searxng
+- ip
+- header
+- configuration
+- proxy_set_header
+- rate
+- logging
+- nginx
+- http_cf_connecting_ip
+- user
+- limiting
+- reverse
+- instance
+links: []
+modified: '2026-04-08T22:26:44+00:00'
+summary: Configuration guide for forwarding real user IPs to SearXNG through Cloudflare
+  Proxy using Nginx CF-Connecting-IP header.
+tags:
+- infrastructure
+- nginx
+- cloudflare
+- searxng
+- proxy
+- security
+title: Cloudflare Proxy and SearXNG IP Header Configuration
+---
+
+# Cloudflare Proxy and SearXNG IP Header Configuration
+
+## Problem
+
+The SearXNG instance behind Cloudflare Proxy only sees requests from the Cloudflare Proxy itself, not the actual IP address of the end user. This causes issues with rate limiting and logging functionality.
+
+## Solution
+
+Every request sent from Cloudflare Proxies includes a separate `CF-Connecting-IP` header containing the real user's IP address. By configuring your reverse proxy (Nginx) to forward this header to SearXNG, you can enable proper IP logging and rate limiting.
+
+## Nginx Configuration
+
+Add the following rules to your Nginx reverse proxy configuration:
+
+```nginx
+proxy_set_header Host $host;
+proxy_set_header Connection $http_connection;
+proxy_set_header X-Forwarded-For $http_cf_connecting_ip;
+proxy_set_header X-Real-IP $http_cf_connecting_ip;
+proxy_set_header X-Scheme $scheme;
+```
+
+## Expected Outcome
+
+- Real user IPs are forwarded to the SearXNG instance
+- IP logging functionality is restored
+- Rate limiting can be re-enabled
+- Should not encounter further rate limit issues
+
+## Additional Use Cases
+
+The `CF-Connecting-IP` header can be useful in other scenarios:
+- Deploying **Fail2Ban** or **CrowdSec** on login prompts
+- Services like **Vaultwarden Server** (some allow header configuration directly)
+- Other reverse proxy setups (Traefik, Caddy alternatives may exist)
+
+## Notes
+
+- SearXNG stores IPs encrypted, so direct verification of logging may not be possible
+- Rate limit issue resolution serves as practical confirmation (95% confidence it works as intended)
+- Alternative to completely disabling Cloudflare proxy (which was recommended but not ideal)
+
+## Related Users
+
+- @unixf0x - Original discovery and recommendation
+- @LincHayes - Also encountered this issue
+
+---
+
+*Source: Cloudflare Proxy SearXNG IP Header Configuration Discussion*