Fine-tuned Llama 3.2 3B to match 70B models for local transcripts
I wrote a small local tool to transcribe audio notes (Whisper/Parakeet). Code: https://github.com/bilawalriaz/lazy-notes I wanted to process raw transcripts locally without OpenRouter. Llama 3.2 3B with a prompt was decent but incomplete, so I tried SFT. I fine-tuned Llama 3.2 3B to clean/analyze dictation and emit structured JSON (title, tags, entities, dates, actions). Data: 13 real memos → Kimi K2 gold JSON → ~40k synthetic + gold; keys canonicalized. Chutes.ai (5k req/day). Training: RTX 4090 24GB, ~4h, LoRA (r=128, α=128, dropout=0.05), max seq 2048,…
In plain words
This tool fine-tunes Llama 3.2 3B to process audio transcripts locally. It transcribes voice notes using Whisper or Parakeet, then cleans and analyzes the text to produce structured JSON output including titles, tags, entities, dates, and action items. The fine-tuned model runs locally via llama.cpp and LM Studio without requiring external APIs. Designed for users who want to process dictation privately on their own hardware.
written from the facts on this page · September 2026
From the sources
In the maker’s words, at launch
I wrote a small local tool to transcribe audio notes (Whisper/Parakeet). Code: https://github.com/bilawalriaz/lazy-notes I wanted to process raw transcripts locally without OpenRouter. Llama 3.2 3B with a prompt was decent but incomplete, so I tried SFT. I fine-tuned Llama 3.2 3B to clean/analyze dictation and emit structured JSON (title, tags, entities, dates, actions). Data: 13 real memos → Kimi K2 gold JSON → ~40k synthetic + gold; keys canonicalized. Chutes.ai (5k req/day). Training: RTX 4090 24GB, ~4h, LoRA (r=128, α=128, dropout=0.05), max seq 2048, bs=16, lr=5e-5, cosine, Unsloth. On 2070 Super 8GB it was ~8h. Inference: merged to GGUF, Q4_K_M (llama.cpp), runs in LM Studio. Evals (100-sample, scored by GLM 4.5 FP8): overall 5.35 (base 3B) → 8.55 (fine-tuned); completeness 4.12 → 7.62; factual 5.24 → 8.57. Head-to-head (10 samples): ~8.40 vs Hermes-70B 8.18, Mistral-Small-24B 7.90, Gemma-3-12B 7.76, Qwen3-14B 7.62. Teacher Kimi K2 ~8.82. Why: task specialization + JSON canonicalization reduces variance; the model learns the exact structure/fields. Lessons: train on completions only; synthetic is fine for narrow tasks; Llama is straightforward to train. Dataset pipeline + training script + evals: https://github.com/bilawalriaz/local-notes-transcribe-llm
Does the same job
all alternatives →- FLFinetune LLaMA-7B on commodity GPUs using your own text2023 · github.com · ▲449
I've been playing around with https://github.com/zphang/minimal-llama/ and https://github.com/tloen/alpaca-lora/blob/main/finetune.py, and wanted to create a simple UI where you can just paste text, tweak the parameters, and finetune the model quickly using a modern GPU. To prepare the data, simply separate your text with two blank lines. There's an inference tab, so you can test how the tuned model behaves. This is my first foray into the world of LLM finetuning, Python, Torch, Transformers, LoRA, PEFT, and Gradio. Enjoy!
- IBI built a free in-browser Llama 3 chatbot powered by WebGPU2024 · github.com · ▲547
I spent the last few days building out a nicer ChatGPT-like interface to use Mistral 7B and Llama 3 fully within a browser (no deps and installs). I’ve used the WebLLM project by MLC AI for a while to interact with LLMs in the browser when handling sensitive data but I found their UI quite lacking for serious use so I built a much better interface around WebLLM. I’ve been using it as a therapist and coach. And it’s wonderful knowing that my personal information never leaves my local computer. Should work on Desktop with Chrome or Edge. Other browsers are adding WebGPU support as well - see…
- LDLlama-dl – high-speed download of LLaMA, Facebook's 65B GPT model2023 · github.com · ▲343
- SHSelf-host Whisper As a Service with GUI and queueing2023 · github.com · ▲267
Schibsted created a transcription service for our journalists to transcribe audio interviews and podcasts really quick.
- NINotesOllama – I added local LLM support to Apple Notes (through Ollama)2024 · smallest.app · ▲156
This lets you talk to local LLMs in Apple Notes. I saw Obsidian Ollama (https://github.com/hinterdupfinger/obsidian-ollama) and thought it was handy, but I'm too lazy to migrate away from the Apple ecosystem, so I quickly hacked this together. I tend to use Notes as a scratchpad for prompts, so it's nice to do some quick inference without leaving the app. Notes doesn't really support plugins so I'm using the macOS accessibility API for reading selections and then stream responses using the clipboard (not ideal but it works).
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 27d ago · cactuscompute.com
Launched alongside, September 2025
the whole month →
- AS
Commerce · Sep 2025 · anycrap.shop
- TE
I made a built-from scratch Wayland Compositor to display any GUI app* in the terminal! I think there is a lot of unexplored potential in custom Wayland compositors, a lot of really cool things you can embed existing applications into! So, I started with embedding apps into the terminal because that is the easiest input/output (output is just utf-8 and I use the great `chafa` library for that, and I just read from stdin for the input). If you have any other ideas for cool Wayland compositors, let me know. I purposedly wrote 80% the app in Typescript to appeal to the most developers and…
Dev tools · Sep 2025 · github.com
- IR
Years ago I stumbled across a basic version of this concept and it stuck with me. I knew if I was ever going to take on such a project, it would need to be flawless, but without coding experience it was just another idea that would never happen. By the end of 2024, as AI coding tools exploded everywhere, I finally had a way to make it real. I started from zero knowledge and spent months collaborating with AI agents as a learning experience. Every pixel and every function went through me. The AI translated what I asked for into code, but every decision was human. I didn't use existing OS…
AI · Sep 2025 · mitchivin.com
