[FEATURE]: integrate DeBERTa-base LocalLLMClient for EDIS fallacy detection

[SoorejS]

Summary

Proposes integrating a fine-tuned DeBERTa-base model as the
LocalLLMClient fallback implementation for EDIS signal extraction.

Why DeBERTa over DistilBERT

DeBERTa-base significantly outperforms DistilBERT on multi-label
fallacy classification due to its disentangled attention mechanism
which handles nuanced reasoning language better.

Model Details

• Base model: microsoft/deberta-base
• Training data: 9,400 debate-style examples
• Fallacy classes: 8
• Validation F1-macro: 1.0
• Hard OOD test: 13/15
• Model: https://huggingface.co/RowdyI7er/DebateLLM

Integration

The model implements the LocalLLMClient interface described
in the EDIS architecture — providing a zero-API-cost fallback
for fallacy detection without any changes to the existing
debate flow or Elo pipeline.

Additional Context

No response

Code of Conduct

• I have joined the Discord server and will post updates there
• I have searched existing issues to avoid duplicates

[SoorejS]

Why the Switch from DistilBERT to DeBERTa-base

The original EDIS proposal described a lightweight DistilBERT
model for local fallacy detection. During training and
evaluation, DistilBERT showed near-random performance on
multi-label fallacy classification — the model was not
capturing the nuanced reasoning structure that separates
fallacy types like Straw Man from Weak Rebuttal.

DeBERTa-base uses a disentangled attention mechanism that
encodes both content and position separately, which makes
it significantly better at understanding the logical
relationships in debate language. After switching to
DeBERTa-base and retraining on 9,400 debate-style examples:

• Validation F1-macro: 1.0
• Hard OOD test (informal language, speed-typed text,
  multi-fallacy sentences): 13/15
• DistilBERT on same data: near-random baseline

This is a better implementation of the same architectural
idea — the LocalLLMClient interface in the EDIS design
remains unchanged. Only the underlying model improves.

[SoorejS]

Model Details

• Base model: microsoft/deberta-base
• Training data: 9,400 debate-style examples
• Fallacy classes: 8 (Ad Hominem, Appeal to Authority, Bandwagon, False Dilemma,
  Hasty Generalization, No Fallacy, Slippery Slope, Strawman)
• Validation F1-macro: 1.0
• Hard OOD test: 13/15
• License: MIT
• Model: https://huggingface.co/RowdyI7er/DebateLLM

Connection to EDIS Proposal

This model directly implements the LocalLLMClient interface
described in Section 6.1 of the EDIS proposal:

• PostDebateAnalyzer interface → model inference endpoint
• FallacySignal.Type → model output class label
• FallacySignal.Phase → passed as context in prompt
• Confidence scoring → model softmax probability output

The switch from DistilBERT to DeBERTa-base does not change
any architectural decision in the EDIS proposal. The
pluggable LLMClient interface was designed exactly for this
— swapping implementations without pipeline changes.

[SoorejS]

References

• EDIS proposal shared with @bhavik-mangla and @bruno in
  AOSSIE Discord, March 25, 2026
• PR chore: add scaffold for post-debate reasoning analysis module #319 — scaffold for post-debate reasoning analysis module
• https://huggingface.co/RowdyI7er/DebateLLM