context
30
ctx:discord/blah/random/30Source document
full textrandom-30
text/plain3 KB
doc:agent/random-30/e62ef2e6-86fc-4562-8c1d-517066f24963[2026-02-17 23:22] xenonfun: mlx fine tuning is dirt simple ``` #!/bin/bash # Fine-tune SmolLM2-360M on SafierSemantics codebase using MLX LoRA # # Usage: ./finetune.sh [iters] # iters: number of training iterations (default: 600) set -e cd "$(dirname "$0")" ITERS=${1:-600} MODEL="HuggingFaceTB/SmolLM2-360M-Instruct" DATA_DIR="./data" ADAPTER_DIR="./adapters" # Step 1: Prepare data if not already done if [ ! -f "$DATA_DIR/train.jsonl" ]; then echo "=== Preparing training data ===" python3 prepare_data.py echo fi # Show data stats echo "=== Training Data ===" echo "Train examples: $(wc -l < "$DATA_DIR/train.jsonl")" echo "Val examples: $(wc -l < "$DATA_DIR/valid.jsonl")" echo "Test examples: $(wc -l < "$DATA_DIR/test.jsonl")" echo # Step 2: LoRA fine-tune echo "=== LoRA Fine-tuning ===" echo "Model: $MODEL" echo "Iterations: $ITERS" echo uvx --from mlx-lm mlx_lm.lora \ --model "$MODEL" \ --train \ --data "$DATA_DIR" \ --iters "$ITERS" \ --batch-size 2 \ --num-layers 8 \ --learning-rate 1e-5 \ --steps-per-report 10 \ --steps-per-eval 50 \ --adapter-path "$ADAPTER_DIR" \ --max-seq-length 2048 echo echo "=== Training complete ===" echo "Adapters saved to: $ADAPTER_DIR/" echo echo "To test: ./test_generate.sh" echo "To fuse into standalone model: ./fuse.sh" ``` (files: Screenshot_2026-02-17_at_6.17.52_PM.png, Screenshot_2026-02-17_at_6.21.53_PM.png) [2026-02-17 23:22] ajaxdavis: lol cool [2026-02-17 23:23] xenonfun: yeah that 360M is about limit of box, said it was only using 6GB but hard to get this thing below 16GB baseline usage, and I have docker and such off. I was bouncing around 22-26gb total so using desktop was not pleasent while training. [2026-02-17 23:25] foxhop.: sounds about right. [2026-02-17 23:25] xenonfun: at least it hasn't cut into my disk too bad yet: ``` SMART/Health Information (NVMe Log 0x02) Critical Warning: 0x00 Temperature: 40 Celsius Available Spare: 100% Available Spare Threshold: 99% Percentage Used: 1% Data Units Read: 206,730,049 [105 TB] Data Units Written: 52,424,513 [26.8 TB] Host Read Commands: 4,767,676,816 Host Write Commands: 1,550,976,917 Controller Busy Time: 0 Power Cycles: 244 Power On Hours: 1,838 Unsafe Shutdowns: 20 Media and Data Integrity Errors: 0 Error Information Log Entries: 0 ``` [2026-02-17 23:26] ajaxdavis: (files: var_02.mp3, var_03.mp3, var_06.mp3, extra_v02_staccato_rock.mp3, extra_v03_dreamy_ambient.mp3, extra_v07_groovy_rnb.mp3) [2026-02-17 23:26] foxhop.: hmmm [2026-02-17 23:27] ajaxdavis: gonna be funny to see the results [2026-02-17 23:28] ajaxdavis: (files: Screenshot_2026-02-18_at_9.28.10_AM.png) [2026-02-17 23:29] ajaxdavis: https://alpha-ashy.vercel.app/chat
Facts in this context
Grouped by subject. Each subject links to its full article.
Smart Nvme Log15 factsex:smart-nvme-log
| hasAttribute | Available Spare |
| hasLogType | NVMe Log 0x02 |
| hasPercentageUsed | 1 |
| rdf:type | Smart Data |
| reportsCriticalWarning | 0x00 |
| reportsDataUnitsRead | 206730049 |
| reportsDataUnitsReadHuman | 105 TB |
| reportsDataUnitsWritten | 52424513 |
| reportsDataUnitsWrittenHuman | 26.8 TB |
| reportsHostReadCommands | 4767676816 |
| reportsHostWriteCommands | 1550976917 |
| reportsPowerCycles | 244 |
| reportsPowerOnHours | 1838 |
| reportsTemperature | 40 |
| reportsUnsafeShutdowns |