khafizovfanis/qwen3.5-4b-discourse-e006-merged

Overview

This model, khafizovfanis/qwen3.5-4b-discourse-e006-merged, is a specialized fine-tuned version of the Qwen3.5-4B base model, with 4.5 billion parameters and a context length of 32768 tokens. It is designed to perform abstract discourse parsing on LLM reasoning traces. The model takes a task prompt and its full reasoning trace (including errors and uncertainties) as input and generates an abstract discourse JSON graph as output. This full BF16 checkpoint has its selected SFT LoRA adapter already merged, requiring no separate PEFT adapter or base model download.

Key Capabilities

• Discourse Parsing: Annotates LLM reasoning traces by generating abstract discourse JSON graphs.
• Error and Uncertainty Handling: Designed to preserve and annotate full reasoning traces, including mistakes, uncertainty, and rejected branches.
• High Schema Validity: Achieved 100% schema-valid outputs on 800 held-out test examples, significantly outperforming the base model (28.125%).
• Specialized Training: Fine-tuned on 9,000 LLM-silver examples with a LoRA adapter (rank 32, alpha 64) over 248 text-backbone Linear modules.

Good for

• LLM Behavior Analysis: Researchers studying the internal reasoning processes of large language models.
• Reinforcement Learning Pipelines: As a component for annotating traces that can then be used to define and test graph-to-reward conversions.
• Trace Annotation: Generating structured representations of complex LLM thought processes for further analysis or visualization.

Limitations

This is a research parser trained on LLM-silver labels, not human gold. Semantic judgments were performed on only 100 out of 800 examples, and no confidence intervals were computed. The model's performance improvement is heavily influenced by a strict failure-zero policy for invalid outputs. It does not establish downstream RL benefit, general-purpose reasoning improvement, or multimodal quality.