Corrected toxicity intepretation based on the new toxicity model beha…

[pjoshi30]

User description

…vior

---

PR Type

Bug fix

---

Description

• Corrected toxicity score interpretation: lower scores now properly indicate higher toxicity

• Inverted confidence calculation (1.0 - score) to accurately reflect toxicity levels

• Updated toxicity threshold from 0.25 to 0.4 with corrected comparison logic

• Enhanced documentation explaining inverted toxicity scoring behavior

---

Diagram Walkthrough

flowchart LR A["Toxicity Detection"] --> B["Score Interpretation"] B --> C["Inverted Score Calculation"] C --> D["Confidence Display"] B --> E["Threshold Comparison"] E --> F["Failure Detection"] 

Loading

File Walkthrough

Relevant files

Bug fix

reprompter.py Invert toxicity confidence calculation for accurate display aimon/reprompting_api/reprompter.py Added documentation note explaining inverted toxicity scoring Changed confidence calculation to (1.0 - score) * 100 to properly reflect toxicity levels +5/-1      utils.py Fix toxicity threshold logic and comparison operators        aimon/reprompting_api/utils.py Updated TOXICITY_THRESHOLD from 0.25 to 0.4 Changed comparison operators from > to < for threshold checks Added documentation clarifying lower scores indicate higher toxicity Updated residual error score calculation comments +13/-10

---

[qodo-merge-pro]

PR Compliance Guide 🔍

Below is a summary of compliance checks for this PR:

Security Compliance
🟢	No security concerns identified No security vulnerabilities detected by AI analysis. Human verification advised for critical code.
Ticket Compliance
⚪	🎫 No ticket provided Create ticket/issue
Codebase Duplication Compliance
⚪	Codebase context is not defined Follow the guide to enable codebase context checks.
Custom Compliance
⚪	No custom compliance provided Follow the guide to enable custom compliance check.
Update

Compliance status legend

🟢 - Fully Compliant
🟡 - Partial Compliant
🔴 - Not Compliant
⚪ - Requires Further Human Verification
🏷️ - Compliance label

[qodo-merge-pro]

PR Code Suggestions ✨

Explore these optional code suggestions:

Category	Suggestion	Impact
High-level	Centralize toxicity score normalization logic To improve maintainability, centralize the toxicity score inversion logic. Create a single normalization function that converts the raw score into a consistent internal representation, preventing the need for widespread changes if the scoring model changes again. Examples: aimon/reprompting_api/utils.py [91] if inst.get("follow_probability", 0.0) < TOXICITY_THRESHOLD aimon/reprompting_api/reprompter.py [99] confidence = (1.0 - failed_instruction.get("score", 0.0)) * 100 Solution Walkthrough: Before: # reprompter.py def get_toxicity_reprompt(result): ... for failed_instruction in failed_instructions: # Logic is inverted here for display confidence confidence = (1.0 - failed_instruction.get("score", 0.0)) * 100 ... # utils.py TOXICITY_THRESHOLD = 0.4 def _count_toxicity_failures(result): # Logic is inverted here for failure check return sum( 1 for inst in ... if inst.get("follow_probability", 0.0) < TOXICITY_THRESHOLD ) After: # In a central utility location def get_normalized_toxicity_score(raw_score): """Returns a normalized score where higher is always more toxic.""" # Current model: lower score = more toxic. We invert it. return 1.0 - raw_score # reprompter.py def get_toxicity_reprompt(result): ... for failed_instruction in failed_instructions: # Use the normalized score directly for confidence normalized_score = get_normalized_toxicity_score(failed_instruction.get("score", 0.0)) confidence = normalized_score * 100 ... # utils.py NORMALIZED_TOXICITY_THRESHOLD = 0.6 # e.g., 1.0 - 0.4 def _count_toxicity_failures(result): # Logic is now consistent (higher > threshold) return sum( 1 for inst in ... if get_normalized_toxicity_score(inst.get("follow_probability", 0.0)) > NORMALIZED_TOXICITY_THRESHOLD ) Suggestion importance[1-10]: 8 __ Why: The suggestion correctly identifies that the logic for handling the inverted toxicity score is scattered, and proposes a valid architectural improvement to centralize it, which would significantly enhance code maintainability and robustness.	Medium
Possible issue	Prevent bugs from missing toxicity scores In get_failed_toxicity_instructions, fetch follow_probability once with a safe default of 1.0 to prevent false positives and improve code clarity. aimon/reprompting_api/utils.py [108-134] def get_failed_toxicity_instructions(result) -> List[dict]: """ Extract failed toxicity instructions below the threshold. Lower scores indicate higher toxicity. Args: result: AIMon detection result containing a `toxicity` section. Returns: List[dict]: A list of dictionaries, each describing a failed toxicity instruction with: - type (str): "toxicity_failure" - source (str): "toxicity" - instruction (str): The instruction text. - score (float): The follow probability. - explanation (str): The explanation for the failure. """ failed = [] for inst in result.detect_response.toxicity.get("instructions_list", []): - if inst.get("follow_probability", 0.0) < TOXICITY_THRESHOLD: + follow_prob = inst.get("follow_probability", 1.0) + if follow_prob < TOXICITY_THRESHOLD: failed.append({ "type": "toxicity_failure", "source": "toxicity", "instruction": inst.get("instruction", ""), - "score": inst.get("follow_probability", 0.0), + "score": follow_prob, "explanation": inst.get("explanation", "") }) return failed Apply / Chat Suggestion importance[1-10]: 8 __ Why: The suggestion correctly identifies a potential bug with missing probabilities and also improves code quality by removing a redundant call to inst.get(), making the code more robust and efficient.	Medium
	Avoid false positives on missing scores In _count_toxicity_failures, change the default for a missing follow_probability from 0.0 to 1.0 to avoid incorrectly flagging it as a toxic failure. aimon/reprompting_api/utils.py [77-92] def _count_toxicity_failures(result) -> int: """ Count the number of toxicity instructions whose follow probability is below the threshold. Lower scores indicate higher toxicity. Args: result: AIMon detection result containing a `toxicity` section. Returns: int: Number of failed toxicity instructions. """ return sum( 1 for inst in result.detect_response.toxicity.get("instructions_list", []) - if inst.get("follow_probability", 0.0) < TOXICITY_THRESHOLD + if inst.get("follow_probability", 1.0) < TOXICITY_THRESHOLD ) Apply / Chat Suggestion importance[1-10]: 7 __ Why: The suggestion correctly identifies a potential bug where a missing follow_probability is treated as a toxicity failure and proposes a valid fix to prevent these false positives.	Medium
Update