Load all JSON files from the directory.

Load all JSON files from the directory. Args: json_dir: Directory containing JSON files offset: Token offset for token-level accu

Load all JSON files from the directory.

Load all JSON files from the directory.

Load all JSON files from the directory.

Load all JSON files from the directory.

Load all JSON results from folder and calculate accuracies keyed by LoRA name.

Load all JSON results from folder and calculate accuracies keyed by LoRA name.

Load all JSON results from folder and calculate accuracies keyed by LoRA name.

Load all JSON files from the directory.

Load Taboo results for Gemma-2-9B-IT, matching the secret-keeping plot logic (taboo_calculate_accuracy with CHOSEN_TABOO_PROMPT and TABOO_SEQ

Load Taboo results for Qwen3-8B (open-ended, one-word secret). Mirrors plot_taboo_eval_results.py for Qwen3-8B with SEQUENCE=False and the

Load all JSON files from the directory.

Load Taboo results from directory - exactly matching original script logic.

Load all JSON files from the directory.

Load only training data from existing dataset files.

Main function to process all JSON files.

Main function to process SAE activations, get explanations, and run precision/recall evaluation. Args: sae_file: Path to the SAE har

Wrap a collator to print tokenization details for debugging.

Create a mask where 1 indicates assistant tokens and 0 indicates non-assistant tokens. Args: tokenized: Dictionary containing 'input

Parse an LLM response containing four Q/A pairs in this form: <question> ... </question> <answer> ... </answer> Returns a list

Create a single plot with all three eval types as subplots.

Create plot with all three models as subplots. Args: all_results: List of result dictionaries for each model model_names: List of

Plot exactly one LoRA (selected by required_keyword in its name) plus extra bars. Asserts that exactly one LoRA matches and that extra_bars h

Plot exactly one LoRA (selected by required_keyword in its name) plus extra bars. Asserts that exactly one LoRA matches and that extra_bars h

Plot exactly one LoRA (selected by required_keyword in its name) plus extra bars. Asserts that exactly one LoRA matches and that extra_bars h

Plot exactly one LoRA (selected by required_keyword in its name) plus extra bars. Asserts that exactly one LoRA matches and that extra_bars h

Plot exactly one LoRA (selected by required_keyword in its name) plus extra bars. Asserts that exactly one LoRA matches and that extra_bars h

Plot exactly one LoRA (selected by required_keyword in its name) plus extra bars. Asserts that exactly one LoRA matches and that extra_bars h

Create separate IID and OOD plots with highlighted LoRA.

Create separate IID and OOD plots with highlighted LoRA.

Create separate plots for each investigator showing per-word accuracy.

Create separate plots for each investigator showing per-word accuracy.

Create separate plots for each investigator showing per-word accuracy.

Create separate plots for each investigator showing per-word accuracy.

Plot one line per series, with: - X-axis = MODEL_TYPE_ORDER - Y-axis = accuracy / quirk score - Color determined by evaluation

Create a bar chart of average accuracy by investigator LoRA.

Create a bar chart of average accuracy by investigator LoRA, highlighting exactly one LoRA.

Create a bar chart of average accuracy by investigator LoRA, highlighting exactly one LoRA.

Create a bar chart of average accuracy by investigator LoRA, highlighting exactly one LoRA.

Create a bar chart of average accuracy by investigator LoRA, highlighting exactly one LoRA.

Create a bar chart of average accuracy by investigator LoRA, highlighting exactly one LoRA.

Create a bar chart of average accuracy by investigator LoRA, highlighting exactly one LoRA.

Create a bar chart of average accuracy by investigator LoRA, highlighting exactly one LoRA.

Create a bar chart of average accuracy by investigator LoRA.

Compute proportion statistics. Returns (p, se, lower, upper) - p: proportion correct (in [0,1]) - se: standard error of the proporti

Read SAEs from a JSONL file with optional start index and limit. Args: sae_file: Path to JSONL file. limit: Maximum number of SAE

Reorder bars: highlight first, then Data Matched, then LatentQA + Classification, then others.

Reorder bars: Full Dataset -> SPQA + Classification -> SPQA Only (Pan et al.) -> Classification -> Original Model.