huggingivy/pragmaprobe-llm-pipeline

System Summary: An end-to-end post-training alignment pipeline for investigating “literal traps” in LLM interpretation of figurative financial language. The system processes financial-news and social-media text archives, including Reuters-derived material, using schema-validated lexicon-guided candidate extraction, MIPVU-informed contextual ranking, and structured JSON mappings. These contextual examples and mappings are converted into a domain-aligned instruction corpus for Low-Rank Adaptation (LoRA) supervised fine-tuning of an open-source Hugging Face transformer model.

To probe the pipeline, selected financial headlines are supplied to both the unadapted base model and the LoRA-adapted model. Their outputs are compared for clarity, contextual grounding, and cross-domain coherence. Where the base model produces an unsupported literal or otherwise ungrounded interpretation, the corresponding curated JSON mapping informs instruction-data construction. The pipeline illustrates how explicit relational mappings can support clearer, contextually grounded figurative-language interpretation within a focused proof-of-concept setting.

Examples of industrial applications for the PragmaProbe-LLM pipeline across key machine learning domains:

• LLM Red Teaming & Behavioral Auditing:
  • Why: Unadapted models may interpret figurative or contextually ambiguous financial language too literally, producing unsupported or poorly grounded explanations rather than the intended business-domain interpretation. Although distinct from this interpretive challenge, LLMs may also be vulnerable to adversarial prompt injections that attempt to redirect instruction following.
  • Application: The pipeline provides a structured post-training workflow for identifying selected contextual interpretation failures and converting curated examples into instruction data for improved figurative-language interpretation.
• Context-Aware Sentiment Recognition:
  • Why: Traditional sentiment analysis tools misclassify intense metaphoric language (e.g., "stocks face a brutal slaughter") as highly violent or negative real-world physical events.
  • Application: A future GraphRAG-enabled extension could retrieve structured contextual mappings for figurative financial expressions, supporting more accurate interpretation and reducing the risk of unsupported literal readings.
• Social Media & Public Relations Management Workflows:
  • Why: Brand-facing consumer bots easily over-generalize or fall into rigid negation loops when responding to colloquial user comments, idioms, or cultural slang.
  • Application: The architecture could be extended for enterprise moderation or customer-service workflows, using automated filtering for routine cases and routing contextually ambiguous or high-impact cases to human review and model-alignment queues.

End-to-End Pipeline Architecture

[Hugging Face: Reuters] ──> [Regex Anchor] ──> [Pandas Staging] ──> [MIPVU-Informed Ranking] ──> [GraphRAG-Inspired JSON Mapping Store] ──> [LoRA / SFT Training] ──> [SFT Model Checkpoint] ──> [Base vs. Adapted Model Probe]

Production Ingestion: Financial-text corpus:

• Retrieves financial-discourse and financial social-media text from the Reuters Financial Corpus through the Hugging Face Datasets API.

• Lexicon-guided candidate extraction: Applies war-domain lexical anchors to identify a curated, lexicon-bounded set of potential metaphor candidates in financial contexts.

• Contextual candidate ranking: Uses MIPVU-informed, context-sensitive ranking to prioritize candidates whose contextual financial meaning differs from a more basic war-domain meaning.

• Typed relational mapping: Links ranked war-domain cues to candidate business and financial meanings trhough JASON-based relational mappings

• LoRA post-training and probing: Converts the structured mappings and contextual examples into an instruction corpus for LoRA-based post-training, then compares base and adapted model interpretations in an illustrative probe.

You can review the full deployment architecture on the PragmaProbe GitHub Repository.