Show HN: Fine-tuned Llama 3.2 3B to match 70B models for local transcripts
Show HN (score: 14)Description
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
More from Show
Show HN: Claudable – OpenSource Lovable that runs locally with Claude Code
Show HN: Claudable – OpenSource Lovable that runs locally with Claude Code Hey, HN! I'm Aaron. I built an open-source Lovable for Claude Code users.<p>Platforms like Lovable, Replit Agent, and Bolt require separate API keys and $25+/month subscriptions. But if you’re already subscribed to Claude Pro or Cursor, you can use those plans directly without extra costs.<p>Claudable runs entirely locally through Claude Code (Cursor CLI also supported) and provides:<p>- Instant UI preview (similar to Lovable)<p>- Web-optimized, production-ready designs<p>- Direct Git integration<p>- One-click Vercel deployment<p>- Zero additional API costs<p>It’s open source and available today. I’m actively developing it and would love community feedback on what features to prioritize next.<p>GitHub: <a href="https://github.com/opactorai/Claudable" rel="nofollow">https://github.com/opactorai/Claudable</a><p>Happy to answer any questions!
Show HN: Memori – Open-Source Memory Engine for AI Agents
Show HN: Memori – Open-Source Memory Engine for AI Agents Hey HN! I'm Arindam, part of the team behind Memori (<a href="https://memori.gibsonai.com/" rel="nofollow">https://memori.gibsonai.com/</a>).<p>Memori adds a stateful memory engine to AI agents, enabling them to stay consistent, recall past work, and improve over time. With Memori, agents don’t lose track of multi-step workflows, repeat tool calls, or forget user preferences. Instead, they build up human-like memory that makes them more reliable and efficient across sessions.<p>We’ve also put together demo apps (a personal diary assistant, a research agent, and a travel planner) so you can see memory in action.<p>Current LLMs are stateless — they forget everything between sessions. This leads to repetitive interactions, wasted tokens, and inconsistent results. When building AI agents, this problem gets even worse: without memory, they can’t recover from failures, coordinate across steps, or apply simple rules like “always write tests.”<p>We realized that for AI agents to work in production, they need memory. That’s why we built Memori.<p>Memori uses a multi-agent architecture to capture conversations, analyze them, and decide which memories to keep active. It supports three modes:<p>- Conscious Mode: short-term memory for recent, essential context. - Auto Mode: dynamic search across long-term memory. - Combined Mode: blends both for fast recall and deep retrieval.<p>Under the hood, Memori is SQL-first. You can use SQLite, PostgreSQL, or MySQL to store memory with built-in full-text search, versioning, and optimization. This makes it simple to deploy, production-ready, and extensible.<p>Memori is backed by GibsonAI’s database infrastructure, which supports:<p>- Instant provisioning - Autoscaling on demand - Database branching & versioning - Query optimization - Point of recovery<p>This means memory isn’t just stored, it’s reliable, efficient, and scales with real-world workloads.<p>We’ve open-sourced Memori under the Apache 2.0 license so anyone can build with it. You can check out the GitHub repo here: <a href="https://github.com/GibsonAI/memori" rel="nofollow">https://github.com/GibsonAI/memori</a>, explore the docs, and join our community on Discord.<p>We’d love to hear your thoughts. Please dive into the code, try out the demos, and share feedback, your input will help shape where we take Memori from here.
Show HN: Ten years of running every day, visualized
Show HN: Ten years of running every day, visualized Today marks ten years, 3653 consecutive days, of running at least one mile every day under the USRSA rules [1]. To celebrate, I built an interactive dashboard that turns a decade of GPX files into charts you can explore.<p>Running has truly changed my life: I've made lifelong friends, explored beautiful places, and more importantly invested into my own health and fitness, which I'm starting to see the positive benefits as I get older.<p>The stack is pretty simple: a NextJS app, with a Postgres database to keep all my running data, and all the stats are pre-computed and cached in Redis, so I effectively only hit the database once a day when a new run is ingested. On the fronted, I toyed with the idea of using D3 or pre-existing data viz libraries, but ended up rolling my own using SVGs directly, it gave me more control on the visualizations.<p>I used the Strava bulk export to pre-populate the database, and I'm using their webhook API to do incremental updates. I have to tap into OpenWeatherMap and OpenCageDate to enrich the running data a little bit.<p>Happy to answer anything about the stack, data pipeline, or how I stayed motivated for 10 years!<p>[1] <a href="https://www.runeveryday.com" rel="nofollow">https://www.runeveryday.com</a> Run Streak Association rules: ≥ 1 mile per day
No other tools from this source yet.