Engine Defaults & Evaluation Presets

This page is a compact, reproducible reference for:

  1. The default engine hyperparameters used by RTSeg when you do not override kwargs.
  2. The default / recommended evaluation settings used by the metrics utilities.

All values below reflect the defaults defined in the code (factory defaults + metric function defaults).

Global RTSeg defaults

When you construct RTSeg(...):

Setting Default Meaning
label_fusion_type "majority" How to fuse labels when multiple engines are used
seg_base_unit "clause" Base unit used by engines that rely on base offsets
model_base "Qwen/Qwen2.5-0.5B-Instruct" Default LLM used by several forced-decoding engines

Engine defaults (factory-provided)

The table below lists the default kwargs injected by RTSeg per engine.

Notes:

  • seg_base_unit is injected into every engine call (whatever you set on RTSeg).
  • Prompt values are stored under keys like system_prompt_offset in your prompt registry; here we reference them by key name (not the full prompt text).
  • You can override any of these at call time: offsets, labels = segmentor(trace, model_name="...", chunk_size=..., ...).
Engine Category Default kwargs (as passed by RTSeg)
RTRuleRegex Rule-based model_name=None, system_prompt=None, seg_base_unit=<RTSeg.seg_base_unit>
RTNewLine Rule-based model_name=None, system_prompt=None, seg_base_unit=<RTSeg.seg_base_unit>
RTLLMOffsetBased LLM (offset boundaries) model_name="Qwen/Qwen2.5-7B-Instruct", system_prompt=load_prompt("system_prompt_offset"), prompt="", chunk_size=300, seg_base_unit=<...>
RTLLMSegUnitBased LLM (segment units) model_name="Qwen/Qwen2.5-7B-Instruct", system_prompt=load_prompt("system_prompt_sentbased"), prompt="", chunk_size=100, seg_base_unit=<...>
RTLLMForcedDecoderBased Forced decoding model_name=<RTSeg.model_base>, system_prompt=load_prompt("system_prompt_forceddecoder"), seg_base_unit=<...>
RTLLMSurprisal Probabilistic (forced decoding) model_name=<RTSeg.model_base>, system_prompt=load_prompt("system_prompt_surprisal"), seg_base_unit=<...> (engine-level window/quantile/max_kv_tokens use the engine’s own defaults unless overridden)
RTLLMEntropy Probabilistic (forced decoding) model_name=<RTSeg.model_base>, system_prompt=load_prompt("system_prompt_surprisal"), seg_base_unit=<...> (engine-level window/quantile/max_kv_tokens use the engine’s own defaults unless overridden)
RTLLMTopKShift Probabilistic (forced decoding) model_name=<RTSeg.model_base>, system_prompt=load_prompt("system_prompt_surprisal"), seg_base_unit=<...> (engine-level top_k/quantile/... use engine defaults unless overridden)
RTLLMFlatnessBreak Probabilistic (forced decoding) model_name=<RTSeg.model_base>, system_prompt=load_prompt("system_prompt_surprisal"), seg_base_unit=<...>
RTBERTopicSegmentation Topic model_name="Qwen/Qwen2.5-1.5B-Instruct", system_prompt=load_prompt("system_prompt_topic_label"), seg_base_unit=<...>
RTZeroShotSeqClassification Zero-shot model_name="facebook/bart-large-mnli", system_prompt="", labels=["verification","pivot","inference","framing","conclusion"], seg_base_unit=<...>
RTZeroShotSeqClassificationRF Zero-shot (RF) model_name="facebook/bart-large-mnli", system_prompt="", seg_base_unit=<...>
RTZeroShotSeqClassificationTA Zero-shot (TA) model_name="facebook/bart-large-mnli", system_prompt="", seg_base_unit=<...>
RTPRMBase PRM-based model_name="Qwen/Qwen2.5-Math-7B-PRM800K", system_prompt="", seg_base_unit=<...>
RTEmbeddingBasedSemanticShift Semantic shift model_name="all-MiniLM-L6-v2", system_prompt="", seg_base_unit=<...>
RTEntailmentBasedSegmentation Entailment / NLI model_name="MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7", system_prompt="", seg_base_unit=<...>
RTLLMThoughtAnchor LLM schema labeling model_name="Qwen/Qwen2.5-7B-Instruct", system_prompt=load_prompt("system_prompt_thought_anchor"), user_prompt=load_prompt("user_prompt_thought_anchor"), seg_base_unit=<...>
RTLLMReasoningFlow LLM schema labeling model_name="Qwen/Qwen2.5-7B-Instruct", system_prompt=load_prompt("system_prompt_reasoning_flow"), user_prompt=load_prompt("user_prompt_reasoning_flow"), seg_base_unit=<...>
RTLLMArgument LLM schema labeling model_name="Qwen/Qwen2.5-7B-Instruct", system_prompt=load_prompt("system_prompt_argument"), user_prompt=load_prompt("user_prompt_argument"), seg_base_unit=<...>

Evaluation settings (defaults + recommended preset)

RT-SEG provides several metric families. The defaults below are important because they directly affect tolerance for “near misses”.

Key parameters

Parameter Unit Default (in metric code) Where it applies
sigma chars 5.0 Soft boundary scoring (Soft_Boundary_F1)
window chars 3 Pairwise Boundary_Similarity in the evaluation registry
slack chars 10 Boundary_Cover (optimistic diagnostic)

Additionally, some helper/aggregate scorers default to:

Function Default Meaning
ReasoningAgreementSuite(window_size=...) 3 Jitter tolerance for boundary similarity (used by the suite)
evaluate_approaches_bounding_similarity(..., window=...) 10 Triadic boundary similarity aggregation default

If you want one consistent, explicit setting to report (and to reproduce tables), a sensible preset that matches your test usage is:

  • sigma=5.0
  • window=3
  • slack=10

When you report results, include these values in the caption/JSON header alongside:

  • dataset/split identifier
  • engine(s) + aligner
  • seg_base_unit
  • model_name and prompt identifiers (for LLM-based engines)
This site uses Just the Docs, a documentation theme for Jekyll.