Simulation: LLMs

Analytical training forecast for language models. Runs SAE gradient decomposition, LiSSA influence scoring, NTK-linearised weight prediction, and dataset quality analysis without real optimizer steps. Produces a synthetic checkpoint you can diff and inspect. Requires LLM mode.

Prerequisiteaquin load --model gpt2-small

6 commands

aquin dataset-generate

agent tool: dataset_generate

Generates instruction/response pairs for a topic and writes a .jsonl file to the current working directory. Used to bootstrap probe datasets before simulation when you do not have training data yet.

Flag	Description
--topic*	Subject matter for generated pairs.
--count	Number of pairs (default: 5).
--output	Output filename (default: <topic>_dataset.jsonl).

example

aquin dataset-analyze

agent tool: analyze_training_dataset

Dataset quality report without loading the model: harmful content scan, AI-fingerprint detection (assistant-speak phrases), length distribution, instruction diversity, and sequence-length violations. Run this before committing GPU time to simulation.

Flag	Description
--dataset	Path to .jsonl / .json / .csv file (default: most recent in cwd).

example

aquin simulate

agent tool: run_simulation

Full LLM simulation pipeline: Pass 0 dataset quality, Pass 1 SAE baseline activations, Pass 2 gradient landscape + SAE gradient decomposition, Pass 2b LiSSA influence, Pass 3 NTK-linearised weight delta. Saves a synthetic checkpoint locally. Takes 2–10 minutes on larger models.

Flag	Description
--dataset	Path to training dataset (.jsonl).
--algo	Path to LoRA/training config YAML.
--topic	Quick probe topic when no dataset yet (auto-generates).
--rank / --lr / --epochs	LoRA hyperparameters (override algo file).
--use_rlhf / --rlhf_beta	Enable RLHF simulation pass.

example

Saved to ~/.aquin/runs/<id>/. Same list/load/compare commands work for LLM and embed runs.

aquin list simulations

agent tool: list_simulation_runs

Lists all saved simulation run IDs on local disk with metadata (topic, model, timestamp).

example

aquin load simulation

agent tool: load_simulation_run

Loads one saved simulation run by ID and renders the full result card on the web.

Flag	Description
--run_id*	Run ID from list simulations.

example

aquin compare simulation

agent tool: compare_simulations

Side-by-side comparison of two saved runs: predicted loss delta, SAE feature diffs, influence score diffs, and attack-surface metrics (consistency, suppression, robustness) from model diff.

Flag	Description
--run_id_a*	First run (before).
--run_id_b*	Second run (after).
--label_a / --label_b	Display labels for the comparison table.

example