🔧 GriceBench-Repair

Rewrites Gricean maxim violations into cooperative dialogue — surgically, not generally.

Part of the GriceBench system — GitHub | 🔍 Detector | ⚡ DPO Generator

---

What This Model Does

GriceBench-Repair is a T5-base seq2seq model that rewrites Gricean maxim violations into cooperative responses. It is violation-type-aware: different maxims use different generation strategies because the nature of the repair task differs.

Violation	Decoding Strategy	Why
Quantity	Beam search (n=4) + length constraints	Needs precise length control
Quality	Beam search (n=4) + repetition penalty	Needs factual precision
Manner	Nucleus sampling (T=0.85, top-p=0.92)	Needs creative diverse rewrites
Relation	NOT this model — use FAISS retrieval	Entire response is off-topic; editing cannot fix it

Violation removal rate: 93.0% (post-fix evaluation, N=200)

---

Quick Start

from transformers import T5ForConditionalGeneration, T5Tokenizer
import torch

model_name = "Pushkar27/GriceBench-Repair"
tokenizer = T5Tokenizer.from_pretrained(model_name)
model = T5ForConditionalGeneration.from_pretrained(model_name)
model.eval()

def repair_violation(context: str, response: str, violation_type: str) -> str:
    assert violation_type in ["quantity", "quality", "manner"], \
        "Relation violations must use the FAISS retrieval system — not this model."

    input_text = f"fix {violation_type}: [CONTEXT] {context} [RESPONSE] {response}"
    inputs = tokenizer(input_text, return_tensors="pt", max_length=256, truncation=True)

    with torch.no_grad():
        if violation_type == "manner":
            output_ids = model.generate(
                **inputs,
                do_sample=True, temperature=0.85, top_p=0.92,
                max_length=128, min_length=8,
                repetition_penalty=1.5, no_repeat_ngram_size=3,
            )
        else:
            output_ids = model.generate(
                **inputs,
                num_beams=4, max_length=128, min_length=8,
                repetition_penalty=1.5, no_repeat_ngram_size=3,
            )

    return tokenizer.decode(output_ids[0], skip_special_tokens=True)

# Quantity (too short)
print(repair_violation(
    context="What do you think about commercial space travel?",
    response="It's fine.",
    violation_type="quantity"
))

# Manner (ambiguous pronouns)
print(repair_violation(
    context="Alice told Bob she would handle the project.",
    response="She said she would do it before she left.",
    violation_type="manner"
))

---

Performance

Violation removal rate: 93.0% (post-fix evaluation)

Per-maxim BLEU scores on the repair validation set (N=401):

Violation Type	BLEU	Notes
Quality	97.8%	Near-perfect factual correction
Manner	92.5%	Strong clarity improvements
Quantity	61.8%	Harder — requires insertions/deletions
Relation	N/A	Route to FAISS retrieval

Degeneracy fix (before vs. after violation-type-aware decoding):

Maxim	Before Fix	After Fix	Improvement
Quantity	30.1% degenerate	2.1%	−28.0pp
Manner	93.3% degenerate	4.5%	−88.8pp
Overall	64.4% degenerate	5.2%	−59.2pp

---

Architecture & Training

• Base model: google-t5/t5-base (220M parameters)
• Training pairs: 3,210 (violation → cooperative) seq2seq pairs
• Validation pairs: 401
• Epochs: 5 | Label smoothing: 0.1 | Hardware: Kaggle T4

Three-layer degeneracy prevention:

1. Violation-type-aware decoding (nucleus sampling for Manner, beam for others)
2. Post-generation multi-signal filter
3. Graceful fallback with is_fallback: True flag

---

Why Relation Violations Use Retrieval

Relation violations mean the entire response is off-topic — there is nothing to edit. We route Relation repairs to a FAISS index over 50,000 Topical-Chat responses (MRR > 0.70, Top-1 accuracy > 60%).

---

Files

File	Description
config.json	T5-base configuration
model.safetensors	Trained model weights
tokenizer.json	SentencePiece tokenizer
tokenizer_config.json	Tokenizer configuration

---

Limitations & Biases

• Hallucination Risk: T5 can occasionally introduce factual errors during repair. Always verify with the "Quality" detector.
• Mode Collapse: Avoid using beam search for "Manner" repairs.

---

Citation

 @article{prabhath2026gricebench,
  title={GriceBench: Operationalizing Gricean Maxims for Cooperative Dialogue Evaluation and Generation},
  author={Prabhath, Pushkar},
  year={2026},
  note={Under review, EMNLP 2026}
}

---

Related Models

Model	Role	Link
GriceBench-Detector	Detects which maxim was violated	🔍 Detector
GriceBench-Repair	Repairs violations (this model)	You are here
GriceBench-DPO	Generates cooperative responses	⚡ DPO

GitHub: https://github.com/PushkarPrabhath27/Research-Model

---

Environmental Impact

Aspect	Value
Hardware Used	NVIDIA Tesla T4 GPU
Training Time	~2 hours
Estimated Carbon Footprint	~0.25 kg CO2eq