A practical guide to sequence packing for ML engineers training LLMs: measuring padding waste and estimating speedup, greedy packing implementation with EOS separation, the attention leakage problem in naive packing, document-aware attention masks with Flash Attention cu_seqlens, TRL SFTTrainer packing configuration, and how to verify packing efficiency and model quality after implementation.
A complete Ollama REST API reference with curl examples for every endpoint: health check, /api/generate with streaming and options, /api/chat with multi-turn history and structured output format parameter, /api/embeddings, /api/tags to list models, /api/pull with progress streaming, /api/delete, /api/copy, /api/create from a Modelfile string, /api/ps for loaded models, /api/show for model details, and the OpenAI-compatible /v1/chat/completions, /v1/models, and /v1/embeddings endpoints.
A practical guide to multi-task learning for ML engineers: hard parameter sharing with task-specific heads, soft parameter sharing with cross-encoder regularisation, gradient cosine similarity for detecting negative transfer, homoscedastic uncertainty loss weighting, task sampling strategies, and an honest assessment of when multi-task training beats separate single-task baselines and when it does not.
A complete guide to Tabby, the open-source self-hosted coding assistant: what it does and how it compares to GitHub Copilot, installing via brew, binary, or Docker, running with built-in code models on CPU and GPU, connecting to VS Code and JetBrains IDEs with API token setup, model selection by hardware tier from 1.3B CPU to 13B GPU, enabling repository context indexing for project-aware completions, and running as a systemd service for persistent availability.
A practical guide to EleutherAI lm-evaluation-harness for ML engineers: CLI and Python API usage, running MMLU, HellaSwag, ARC and TruthfulQA, evaluating fine-tuned checkpoints, writing custom YAML tasks for domain benchmarks, understanding acc vs acc_norm vs mc2 metrics, and avoiding the prompt format mismatches and contamination issues that produce misleading benchmark numbers.
A practical guide to Mistral Nemo 12B: its distinctive technical features including 128K native context, the Tekken tokeniser, and strong multilingual training across 11 languages, hardware requirements at Q4_K_M (~7GB), when the 12B quality jump over 7–8B models is worth the extra VRAM, a Modelfile for long-context use, multilingual Python examples, and a clear comparison against Llama 3.2 8B and where Nemo wins.
A practical guide to text data augmentation for ML engineers: why text augmentation is harder than image augmentation, word-level perturbations with EDA, back-translation with MarianMT, LLM-based paraphrasing for instruction datasets, embedding-space Mixup for classification, and how to verify empirically that augmentation is actually helping rather than hurting model quality.
A complete guide to the local audio summarisation pipeline using faster-whisper for transcription and Ollama for summarisation: installing faster-whisper and ffmpeg, transcribing with different model sizes from tiny to large-v3, a full pipeline function that handles long transcripts by chunking, four summary styles (bullets, paragraph, tldr, chapters), extracting action items and decisions from meeting recordings, and a command-line script for quick use from the terminal.
A practical guide to PyTorch Lightning for ML engineers working on LLMs: LightningModule structure for fine-tuning, LightningDataModule for reproducible data pipelines, Trainer configuration for multi-GPU with FSDP and mixed precision, LoRA fine-tuning with PEFT integration, gradient accumulation and checkpointing, and a clear-eyed comparison of when Lightning helps versus when Accelerate or raw PyTorch DDP is the better choice.
A practical guide to running Llama 3.3 70B locally with Ollama: hardware requirements across Apple Silicon configurations and NVIDIA GPU setups, pulling the model and verifying GPU layer loading with ollama ps, configuring large context windows with a Modelfile, the four specific task areas where 70B quality significantly outperforms 7-8B models, Python usage for complex reasoning tasks with streaming, realistic tokens per second on M3 Max and dual RTX 4090, and a decision framework for when to use 70B versus smaller models.
