tyrin/preference-judge
1
1
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

(no commit message)

Browse Source
This commit is contained in:
Tyrin Todd
2026-01-24 20:08:17 -08:00
parent 942d32f9e5
commit 649667e2e7
3 changed files with 428 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

325
README.md
View File

@@ -1,2 +1,325 @@
# preference-judge # Bench
Modaic internal SDK for benchmarking judges and training confidence probes.
## Installation
```bash
uv sync
```
## CLI Commands
All commands can be run via `uv run bench <command>` or using the shorthand `uv run <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:**
```bash
# Interactive mode (recommended) - prompts for configuration
uv run create ppe
uv run create judge_bench
# With config file
uv run create ppe --config config.yaml
uv run create judge_bench --config config.yaml
# Full command form
uv run bench create ppe --config config.yaml
```
**Options:**
| Option | Short | Description |
| ---------- | ----- | -------------------------- |
| `--config` | `-c` | Path to config file (YAML) |
**Config File Example:**
```yaml
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:**
```bash
# Interactive mode (recommended) - prompts for all configuration
uv run train
# With config file
uv run train --config config.yaml
# With CLI arguments
uv run train --dataset tytodd/my-embeddings --epochs 10 --lr 0.0001
# Full command form
uv run bench train --config config.yaml
```
**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:**
```yaml
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:**
```bash
# Interactive mode (recommended) - prompts for probe and dataset
uv run eval
# With CLI arguments
uv run eval --probe tytodd/my-probe --dataset tytodd/my-embeddings
# Evaluate on train split instead of test
uv run eval --probe tytodd/my-probe --dataset tytodd/my-embeddings --split train
# Full command form
uv run bench eval --probe tytodd/my-probe --dataset tytodd/my-embeddings
```
**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:**
```bash
# Interactive mode
uv run compile
uv run compile ppe
# With config file
uv run compile --config config.yaml
uv run compile ppe --config config.yaml
# Full command form
uv run bench compile --config config.yaml
```
**Options:**
| Option | Short | Description |
| ---------- | ----- | -------------------------- |
| `--config` | `-c` | Path to config file (YAML) |
**Config File Example:**
```yaml
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
---
### `reembed`
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:**
```bash
# Interactive mode
uv run reembed
# With CLI arguments
uv run reembed --dataset tytodd/my-dataset --hf-model Qwen/Qwen3-VL-32B-Instruct --layer -1
# Full command form
uv run bench reembed --dataset tytodd/my-dataset
```
**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:**
```bash
# Original dataset was created with layer 32
# Now try middle layer instead
uv run reembed \
--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
```bash
# 1. Create a probe dataset from a benchmark
uv run create ppe
# 2. Train a confidence probe
uv run train --dataset tytodd/ppe-qwen3-embeddings
# 3. Evaluate the probe on a test set
uv run eval --probe tytodd/my-probe --dataset tytodd/ppe-qwen3-embeddings
# 4. (Optional) Compile/optimize a judge with GEPA
uv run compile ppe
# 5. (Optional) Re-embed with different layer
uv run reembed --dataset tytodd/my-dataset --layer 32
```
## Environment Variables
Create a `.env` file with:
```bash
OPENAI_API_KEY=...
WANDB_API_KEY=...
HF_TOKEN=...
MODAIC_API_KEY=...
```

56
config.json Normal file
View File

@@ -0,0 +1,56 @@
{
"model": null,
"signature": {
"description": "Evaluate and compare the quality of two responses (Response A and Response B) given a specific question.\nDetermine which response better addresses the question by focusing on factual correctness, completeness,\nand adherence to any specific requirements mentioned in the question prompt.\n\nBefore yielding your decision, think step by step and explain your reasoning in the reasoning field.\nBe sure to verbally express your uncertainty in your thought process.\n\nDetailed Instructions:\n\n1. **Understand the Question Context:**\n - Ensure you comprehend the full context and requirements specified by the question or problem statement.\n - Note any domain-specific terminologies or conditions.\n\n2. **Evaluate Each Response:**\n - Check for factual accuracy in the content, calculations, or recommendations provided.\n - Assess the response for completeness\u2014whether it completely addresses all aspects of the question.\n - Verify adherence to the specified question requirements.\n - Consider clarity and structure of the explanation or solution provided.\n\n3. **Decision Making:**\n - Determine which response (A or B) best meets the above criteria.\n - Select the response that is not only correct but also most aligns with the question's specific requirements.\n\n4. **Output Your Conclusion:**\n - Document your reasoning process in the reasoning field.\n - Output \"A>B\" if Response A is better, or \"B>A\" if Response B is better.",
"properties": {
"question": {
"__dspy_field_type": "input",
"desc": "The original question or prompt",
"prefix": "Question:",
"title": "Question",
"type": "string"
},
"response_A": {
"__dspy_field_type": "input",
"desc": "First response to evaluate",
"prefix": "Response A:",
"title": "Response A",
"type": "string"
},
"response_B": {
"__dspy_field_type": "input",
"desc": "Second response to evaluate",
"prefix": "Response B:",
"title": "Response B",
"type": "string"
},
"reasoning": {
"__dspy_field_type": "output",
"desc": "Your step by step reasoning for why you chose the better response. With verbally expressed uncertainty.",
"prefix": "Reasoning:",
"title": "Reasoning",
"type": "string"
},
"label": {
"__dspy_field_type": "output",
"desc": "Which response is better: 'A>B' or 'B>A'",
"enum": [
"A>B",
"B>A"
],
"prefix": "Label:",
"title": "Label",
"type": "string"
}
},
"required": [
"question",
"response_A",
"response_B",
"reasoning",
"label"
],
"title": "PreferenceSig",
"type": "object"
}
}

48
program.json Normal file
View File

@@ -0,0 +1,48 @@
{
"traces": [],
"train": [],
"demos": [],
"signature": {
"instructions": "Evaluate and compare the quality of two responses (Response A and Response B) given a specific question.\nDetermine which response better addresses the question by focusing on factual correctness, completeness,\nand adherence to any specific requirements mentioned in the question prompt.\n\nBefore yielding your decision, think step by step and explain your reasoning in the reasoning field.\nBe sure to verbally express your uncertainty in your thought process.\n\nDetailed Instructions:\n\n1. **Understand the Question Context:**\n - Ensure you comprehend the full context and requirements specified by the question or problem statement.\n - Note any domain-specific terminologies or conditions.\n\n2. **Evaluate Each Response:**\n - Check for factual accuracy in the content, calculations, or recommendations provided.\n - Assess the response for completeness\u2014whether it completely addresses all aspects of the question.\n - Verify adherence to the specified question requirements.\n - Consider clarity and structure of the explanation or solution provided.\n\n3. **Decision Making:**\n - Determine which response (A or B) best meets the above criteria.\n - Select the response that is not only correct but also most aligns with the question's specific requirements.\n\n4. **Output Your Conclusion:**\n - Document your reasoning process in the reasoning field.\n - Output \"A>B\" if Response A is better, or \"B>A\" if Response B is better.",
"fields": [
{
"prefix": "Question:",
"description": "The original question or prompt"
},
{
"prefix": "Response A:",
"description": "First response to evaluate"
},
{
"prefix": "Response B:",
"description": "Second response to evaluate"
},
{
"prefix": "Reasoning:",
"description": "Your step by step reasoning for why you chose the better response. With verbally expressed uncertainty."
},
{
"prefix": "Label:",
"description": "Which response is better: 'A>B' or 'B>A'"
}
]
},
"lm": {
"model": "together_ai/Qwen/Qwen3-VL-32B-Instruct",
"model_type": "chat",
"cache": true,
"num_retries": 3,
"finetuning_model": null,
"launch_kwargs": {},
"train_kwargs": {},
"temperature": null,
"max_tokens": null
},
"metadata": {
"dependency_versions": {
"python": "3.11",
"dspy": "3.1.0",
"cloudpickle": "3.1"
}
}
}