mirror of
https://github.com/p-e-w/heretic.git
synced 2026-11-06 05:16:39 +00:00
Vinay Umrethe
ecde93f1e0
* feat: Refusal classifier scorer plugin Using the generic BERTClassifier scorer plugin * docs: Add modified_model_type * fix: Close [red] tag in the error
121 lines
4.7 KiB
TOML
121 lines
4.7 KiB
TOML
# Rename this file to config.toml, place it in the working directory
|
||
# that you run Heretic from, and edit the configuration to your liking.
|
||
#
|
||
# BERTClassifier scorer for personality/emotion steering.
|
||
#
|
||
# This uses a DistilRoBERTa model fine-tuned on emotion classification
|
||
# and allows minimizing or maximizing specific emotions in the model's responses.
|
||
#
|
||
# Example use: configure two scorer instances, one to minimize negative emotions
|
||
# (e.g. anger, disgust, fear, sadness) and one to maximize positive emotions
|
||
# (e.g joy, surprise) or vice versa.
|
||
#
|
||
# This can be done in one of the two ways:
|
||
#
|
||
# 1. Multi‑objective steering:
|
||
# Here, we use two instances: `negative_emotions` and `positive_emotions` with
|
||
# `optimization = "minimize"` and `optimization = "maximize"` respectively.
|
||
#
|
||
# Minimizing the first score pushes the model *away* from those negative emotions.
|
||
# Maximizing the second score pushes the model *towards* those emotions.
|
||
#
|
||
# 2. Single‑objective steering:
|
||
# With two instances, heretic builds a Pareto front that **balances** both objectives above.
|
||
# Due to which, the model learns to be less angry *and* more joyful **simultaneously**.
|
||
#
|
||
# With only one instance, we have only one objective which can be either minimized or maximized.
|
||
# Due to which, the model will reduce anger, but it *may* drift towards **anything** that lowers
|
||
# the other most natural instincts (e.g. neutral, overly polite, or wierdly happy).
|
||
#
|
||
# With the first method, you can expect to control both objectives effectively,
|
||
# which is the recommended method and is the default.
|
||
# With the second method, you can only *hope* and *assume* the most expected common/natural
|
||
# behavior, depending upon what you tried to "minimize" or "maximize".
|
||
scorers = [
|
||
{ plugin = "heretic.scorers.kl_divergence.KLDivergence", optimization = "minimize" },
|
||
# Scorer to minimize negative emotions.
|
||
{ plugin = "heretic.scorers.bert_classifier.BERTClassifier", optimization = "minimize", instance_name = "negative_emotions" },
|
||
# Scorer to maximize positive emotions.
|
||
{ plugin = "heretic.scorers.bert_classifier.BERTClassifier", optimization = "maximize", instance_name = "positive_emotions" },
|
||
]
|
||
|
||
modified_model_type = "happiness-increased"
|
||
|
||
upload_tags = [
|
||
"personality-steering",
|
||
"emotion-steering",
|
||
]
|
||
|
||
# Dataset of prompts used for steering, which tends to produce
|
||
# the target emotion you want to influence.
|
||
# For example:
|
||
# - For minimizing anger: prompts about creative writing which involves those emotions.
|
||
# - For maximizing joy: prompts about celebrations, happy scenarios, good news from user, etc.
|
||
[modifier.ARA.good_prompts]
|
||
dataset = "VINAY-UMRETHE/Positive-Prompts"
|
||
split = "train[:200]"
|
||
column = "prompt"
|
||
|
||
[modifier.ARA.bad_prompts]
|
||
dataset = "VINAY-UMRETHE/Negative-Prompts"
|
||
split = "train[:200]"
|
||
column = "prompt"
|
||
|
||
|
||
[scorer.BERTClassifier_negative_emotions]
|
||
# Name that describes what the configured match rate measures.
|
||
score_name = "Minimize"
|
||
|
||
# Hugging Face model ID, or path to a text-classification model on disk.
|
||
model = "j-hartmann/emotion-english-distilroberta-base"
|
||
|
||
# Hugging Face commit hash of the classifier model.
|
||
# Set this to pin a specific revision; unset means the latest revision
|
||
# is auto-pinned and recorded when the classifier is loaded.
|
||
# When reproducing, the recorded revision is restored and used.
|
||
# model_commit = "<commit-hash>"
|
||
|
||
# Labels predicted by the classifier that identify a response as a match.
|
||
# We want to minimize these emotions.
|
||
match_labels = ["anger", "disgust", "fear", "sadness"]
|
||
|
||
# Number of prompt/response pairs to classify in parallel.
|
||
batch_size = 16
|
||
|
||
# Maximum number of tokens per classifier input (longer inputs are truncated).
|
||
max_length = 512
|
||
|
||
# Whether to pass inputs as prompt/response text pairs.
|
||
# Requires a model trained on pairs (e.g. the default refusal classifier).
|
||
# Set to False when using a single-sequence text-classification model
|
||
# to evaluate the response string only (e.g. sentiment analysis).
|
||
use_text_pairs = false
|
||
|
||
# Whether to print prompt/response pairs and their predicted labels.
|
||
print_responses = true
|
||
|
||
# Dataset of prompts to evaluate the match rate on.
|
||
# Use a dataset that tends to produce the target emotions (anger, disgust, fear, sadness).
|
||
[scorer.BERTClassifier_negative_emotions.prompts]
|
||
dataset = "VINAY-UMRETHE/Negative-Prompts"
|
||
split = "train[:100]"
|
||
column = "prompt"
|
||
|
||
|
||
[scorer.BERTClassifier_positive_emotions]
|
||
score_name = "Maximize"
|
||
|
||
model = "j-hartmann/emotion-english-distilroberta-base"
|
||
|
||
# We want to maximize these emotions.
|
||
match_labels = ["joy", "surprise"]
|
||
|
||
batch_size = 16
|
||
max_length = 512
|
||
print_responses = true
|
||
|
||
[scorer.BERTClassifier_positive_emotions.prompts]
|
||
dataset = "VINAY-UMRETHE/Positive-Prompts"
|
||
split = "train[:100]"
|
||
column = "prompt"
|