tri train — HSLM Training Monitor

Monitor and control HSLM (Hyper-Sparse Language Model) training runs. Supports both local and Railway-hosted training.

Subcommands

Command	Arguments	Description
tri train status	[--json] [--host railway]	Live dashboard or JSON status
tri train start	[options] [--host railway]	Launch training (local or remote)
tri train logs	[--host railway]	Tail training logs
tri train loss [dir]	[checkpoint-dir]	Parse checkpoint loss curve
tri train diagnose [dir]	[checkpoint-dir]	Auto-diagnose training anomalies
tri train compare <d1> <d2>	<dir1> <dir2>	Side-by-side run comparison
tri train checkpoint list [dir]	[checkpoint-dir]	List checkpoints with metrics

Options for tri train start

Option	Default	Description
--steps <N>	100000	Total training steps
--lr <value>	3e-4 (local), 1e-4 (railway)	Learning rate
--warmup <N>	5000	Warmup steps
--batch <N>	64	Batch size
--optimizer <type>	adamw	Optimizer: adamw/lamb
--ste <mode>	none	STE mode: none/vanilla/twn/progressive
--wd <value>	0.1	Weight decay
--checkpoint-dir <path>	data/checkpoints	Checkpoint directory
--resume <path>	—	Resume from checkpoint
--data <path>	data/tinystories/real_tinystories.txt	Training data file
--grad-accum <N>	1	Gradient accumulation steps
--context <N>	81	Context length

Examples

tri train status                   # Live training dashboard
tri train start --lr 1e-4          # Start local training
tri train start --host railway     # Start Railway training
tri train loss data/checkpoints    # Show loss curve
tri train diagnose data/checkpoints # Diagnose issues
tri train compare run1/ run2/      # Compare two runs
tri train checkpoint list          # List saved checkpoints

Environment Variables

Variable	Required	Description
HSLM_OPTIMIZER	Yes (remote)	Optimizer type
HSLM_LR	Yes (remote)	Learning rate
HSLM_LR_SCHEDULE	Yes (remote)	LR schedule (ALWAYS cosine)

Handler

File: src/tri/tri_train.zig