hgLeo12/RoTRAG_roberta_routing_classifier

RoTRAG RoBERTa Routing Classifier

This model is the routing classifier used in RoTRAG: Retrieval-Augmented Rule-of-Thumb Reasoning for Conversation Harm Detection.

The classifier decides whether a previously generated Rule of Thumb (RoT) can be carried over to the current dialogue turn or whether RoTRAG should generate a new RoT for the current context.

Label Space

Label	Meaning
0	The previous RoT is no longer sufficient; generate a new RoT.
1	The previous RoT can be retained for the current turn.

Intended Use

This checkpoint is intended for the router step in the RoTRAG pipeline. Given the previous RoT information and the current context-response pair, it predicts whether the normative core of the previous RoT still applies.

It is not a standalone harm detector. The classifier only decides RoT carry-over versus RoT regeneration; downstream safety or prosocial labels are produced by separate RoTRAG reasoning and prediction steps.

Model Details

• Architecture: RobertaForSequenceClassification
• Base model: roberta-large
• Task: binary sequence classification
• Input type: previous RoT(s), current dialogue context, and current response
• Output: binary routing label (0 or 1)

Training Data

The routing supervision data was constructed from ProsocialDialog train/validation instances. The annotation process first used 1,000 human-labeled routing examples, annotated by 10 expert annotators. The final gold label for each example was determined by hard voting.

The resulting annotation rule was distilled into a prompt and validated against the human-labeled subset, reaching 0.98 accuracy relative to the human gold labels. The validated prompt was then used to expand labels over the remaining training data.

Expanded label counts reported for classifier training:

Split	Label 0	Label 1
Train	14,528	46,206
Validation	2,489	8,147

Training Setup

The classifier was fine-tuned with the following setup:

Hyperparameter	Value
Base model	roberta-large
Epochs	3
Learning rate	2e-5
Scheduler	Linear, no warm-up
Optimizer	AdamW
Weight decay	0.01
Train batch size	8
Gradient accumulation	2
Effective batch size	16
Evaluation batch size	16
Max sequence length	256
Padding	Dynamic
Mixed precision	FP16
Best checkpoint criterion	Macro F1
Random seed	42

Evaluation

On 10,636 validation routing instances, the classifier produced the following confusion matrix counts:

Outcome	Count	Share
True positive	7,461	70.215%
True negative	1,743	16.403%
False positive	746	7.021%
False negative	676	6.362%

The reported behavior is strongest when the current turn clearly preserves or clearly departs from the previous RoT. Remaining errors are often borderline cases where the current response is partially related to the previous RoT but shifts specificity, framing, or normative emphasis.

Example Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model_id = "hgLeo12/roberta_classifier"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

previous_rot = "You should not manipulate others."
current_context = "She gets what she gets. Her bad habits are getting out of control."
current_response = (
    "You would have more luck if you just talked to her and explained the problem. "
    "If she does not want to change, then maybe you should move on."
)

text = (
    f"Previous RoT: {previous_rot}\n"
    f"Current Context: {current_context}\n"
    f"Current Response: {current_response}"
)

inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)

with torch.no_grad():
    logits = model(**inputs).logits

predicted_label = int(logits.argmax(dim=-1).item())
print(predicted_label)

Interpretation:

• 0: generate a new RoT.
• 1: carry over the previous RoT.

Limitations

• The model is trained for RoT routing, not final safety classification.
• The decision boundary can be ambiguous when the current response partially preserves the previous RoT while changing its framing or specificity.
• Performance should be interpreted in the context of the RoTRAG pipeline and the ProsocialDialog-derived supervision setup.

Related Repository

GitHub repository: GitLeo1/RoTRAG-Rule-of-Thumb-Reasoning-for-Conversation-Harm-Detection-with-Retrieval-Augmented-Generation