Bench

Modaic internal SDK for benchmarking judges and training confidence probes.

Installation

cd cli
uv sync

CLI Commands

All commands are run from the cli directory via uv run mo <command>.

create

Create benchmark datasets for training confidence probes. This command runs a judge on examples, extracts embeddings via Modal, and pushes the resulting dataset to HuggingFace Hub.

Subcommands:

• create ppe - Create dataset from PPE (human-preference + correctness) benchmarks
• create judge_bench - Create dataset from the JudgeBench benchmark

Usage:

# Interactive mode (recommended) - prompts for configuration
uv run mo create ppe
uv run mo create judge_bench

# With config file
uv run mo create ppe --config config.yaml
uv run mo create judge_bench --config config.yaml

Options:

Option	Short	Description
--config	-c	Path to config file (YAML)

Config File Example:

judge: tyrin/ppe-judge-gepa
output: tytodd/my-probe-dataset
n_train: 500
n_test: 100
embedding_layer: -1  # -1 for middle layer

What it does:

1. Loads examples from the benchmark dataset
2. Runs the specified judge on each example to get predictions
3. Extracts embeddings from the judge's LLM via Modal (GPU)
4. Creates a HuggingFace dataset with columns: question, response_a, response_b, label, predicted, messages, embeddings
5. Pushes to HuggingFace Hub

---

train

Train a confidence probe on an embeddings dataset created with create.

Usage:

# Interactive mode (recommended) - prompts for all configuration
uv run mo train

# With config file
uv run mo train --config config.yaml

# With CLI arguments
uv run mo train --dataset tytodd/my-embeddings --epochs 10 --lr 0.0001

Options:

Option	Short	Description	Default
--config	-c	Path to config file (YAML)	-
--dataset	-d	Dataset path (HuggingFace Hub or local) (must be a dataset created with create)	-
--model-path	-m	Output path for trained model	{dataset}_probe
--batch-size		Batch size	4
--epochs		Number of training epochs	10
--lr		Learning rate	0.0001
--weight-decay		Weight decay	0.01
--test-size		Validation split ratio (if no test split)	0.2
--seed		Random seed	42
--project		W&B project name	model_path
--hub-path		HuggingFace Hub path to push model	-

Config File Example:

dataset_path: tytodd/my-probe-dataset
model_path: ./best_probe
hub_path: tytodd/my-probe  # Optional: push to HF Hub
batch_size: 4
epochs: 10
learning_rate: 0.0001
weight_decay: 0.01
test_size: 0.2
seed: 42

What it does:

1. Loads an embeddings dataset (from HuggingFace Hub or local)
2. Creates binary labels: 1 if predicted == label, 0 otherwise
3. Trains a linear probe using MSE loss (Brier score optimization)
4. Logs metrics to Weights & Biases (Brier, ECE, MCE, Kuiper, AUROC)
5. Saves the best model based on validation Brier score
6. Optionally pushes to HuggingFace Hub

---

eval

Evaluate a trained confidence probe on a dataset. Computes calibration and discrimination metrics.

Usage:

# Interactive mode (recommended) - prompts for probe and dataset
uv run mo eval

# With CLI arguments
uv run mo eval --probe tytodd/my-probe --dataset tytodd/my-embeddings

# Evaluate on train split instead of test
uv run mo eval --probe tytodd/my-probe --dataset tytodd/my-embeddings --split train

Options:

Option	Short	Description	Default
--probe	-p	Probe path (HuggingFace Hub or local)	-
--dataset	-d	Dataset path (HuggingFace Hub or local)	-
--split	-s	Dataset split to evaluate on	test
--batch-size	-b	Batch size for evaluation	64
--normalize/--no-normalize	-n	Normalize embeddings with StandardScaler	probe config

Metrics computed:

Metric	Description
Brier Score	Mean squared error between predictions and labels
Accuracy	Classification accuracy at 0.5 threshold
F1 Score	Harmonic mean of precision and recall
ECE	Expected Calibration Error (10 bins)
MCE	Maximum Calibration Error
Kuiper	Kuiper statistic for calibration
AUROC	Area Under the ROC Curve (discrimination)

What it does:

1. Loads a pretrained probe from HuggingFace Hub or local path
2. Loads a dataset created with create
3. Creates binary labels: 1 if predicted == label, 0 otherwise
4. Runs inference and computes calibration/discrimination metrics
5. Displays results in a formatted table

---

compile

Compile (optimize) a judge using GEPA over a dataset. GEPA iteratively improves the judge's prompt based on training examples.

Subcommands:

• compile (base) - Compile with custom dataset and parameter mapping
• compile ppe - Compile specifically for PPE datasets (human-preference + correctness)

Usage:

# Interactive mode
uv run mo compile
uv run mo compile ppe

# With config file
uv run mo compile --config config.yaml
uv run mo compile ppe --config config.yaml

Options:

Option	Short	Description
--config	-c	Path to config file (YAML)

Config File Example:

judge: tyrin/ppe-judge
dataset: tytodd/ppe-human-preference
inputs: # selects which input columns of the dataset to use (not necearry if using a compile subcommand like ppe or judge_bench) 
  - name: question
  - name: response_a
    column: response_A  # Map param name to dataset column
  - name: response_b
    column: response_B
label_column: label
n_train: 100
n_val: 50
base_model: gpt-4o-mini
reflection_model: gpt-4o
output: tyrin/ppe-judge-gepa
seed: 42

What it does:

1. Loads a judge from Modaic Hub
2. Loads training/validation examples from a HuggingFace dataset
3. Maps judge parameters to dataset columns
4. Runs GEPA optimization to improve the judge's prompt
5. Pushes the optimized judge to Modaic Hub

---

embed

Regenerate embeddings for an existing dataset using a different model or layer. Useful for experimenting with different embedding configurations without re-running the judge.

Usage:

# Interactive mode
uv run mo embed

# With CLI arguments
uv run mo embed --dataset tytodd/my-dataset --hf-model Qwen/Qwen3-VL-32B-Instruct --layer -1

Options:

Option	Short	Description
--dataset	-d	Dataset path (HuggingFace Hub or local)
--hf-model	-m	HuggingFace model path for embeddings
--layer	-l	Hidden layer index (-1 for middle layer)

What it does:

1. Loads an existing dataset (must have a messages column)
2. Regenerates embeddings using the specified model/layer via Modal
3. Replaces the embeddings column in the dataset
4. Prompts to push the updated dataset to HuggingFace Hub

Example workflow:

# Original dataset was created with layer 32
# Now try middle layer instead
uv run mo embed \
  --dataset tytodd/my-embeddings \
  --hf-model Qwen/Qwen3-VL-32B-Instruct \
  --layer -1

---

Recommended Embedding Layers

When extracting embeddings, use these recommended layer indices for best probe performance:

Model	HuggingFace Path	Recommended Layer
GPT-OSS 20B	openai/gpt-oss-20b	8
Qwen3-VL 32B	Qwen/Qwen3-VL-32B-Instruct	16
Llama 3.3 70B	meta-llama/Llama-3.3-70B-Instruct	32

Use -1 for the middle layer if experimenting with an unlisted model.

---

Typical Workflow

# 1. Create a probe dataset from a benchmark
uv run mo create ppe

# 2. Train a confidence probe
uv run mo train --dataset tytodd/ppe-qwen3-embeddings

# 3. Evaluate the probe on a test set
uv run mo eval --probe tytodd/my-probe --dataset tytodd/ppe-qwen3-embeddings

# 4. (Optional) Compile/optimize a judge with GEPA
uv run mo compile ppe

# 5. (Optional) Re-embed with different layer
uv run mo embed --dataset tytodd/my-dataset --layer 32

Environment Variables

Create a .env file with:

OPENAI_API_KEY=...
WANDB_API_KEY=...
HF_TOKEN=...
MODAIC_TOKEN=...
TOGETHER_API_KEY=...