TBG-CoT-Bench

TBG-CoT-Bench is a local application benchmark for testing temporal belief tracking over Chain-of-Thought-style evidence sequences.

The benchmark evaluates whether a system can track belief about the temporal claim:

Event A occurred before Event B.

This repository contains synthetic temporal reasoning scenarios, rule-based baselines, local EXAONE/Ollama experiments, trajectory visualizations, generated reports, and pytest-based application benchmark checks.

---

Current Status

Application benchmark tests: 13 passed
Best current method: EXAONE step-wise evidence v2.1

The current best local method is EXAONE step-wise evidence extraction v2.1, which achieved stronger direction accuracy than the rule-based converter baseline while maintaining high structured-output parse stability.

---

Main Results

Method	Parse success rate	Direction accuracy	Notes
Rule-based baseline converter	N/A	0.5769	Deterministic converter using hand-written extraction rules.
EXAONE scenario-level CoT	Low / unstable	0.4615 on parsed steps	Full-scenario prompting produced frequent JSON parse failures.
EXAONE step-wise evidence v2.1	0.9615	0.6600	Best current local method; 50 / 52 steps parsed.
EXAONE order-classification v3	0.6154	Unstable	High UNCLEAR collapse; directional accuracy is not directly comparable due to few directional parses.
EXAONE cumulative belief v4	0.4423	Weak trajectory agreement	Conceptually close to belief tracking but less stable in the local setup.

Main interpretation: step-wise structured evidence extraction is currently the most reliable local architecture for converting CoT-style temporal evidence into belief trajectories.

---

Key Visualizations

Stepwise EXAONE/Ollama vs Baseline

Stepwise Parse Success

Gold Belief Trajectories

Stepwise EXAONE/Ollama Trajectories

---

Scenario-Level Visualizations

Scenario	Stepwise trajectory comparison
Scenario 01
Scenario 02
Scenario 03
Scenario 04
Scenario 05
Scenario 06
Scenario 07
Scenario 08
Scenario 09
Scenario 10

---

What This Repository Tests

This repository is an application-level benchmark / usage test, not a full internal unit test suite for the upstream temporal-belief-graph package.

It tests the following workflow:

scenario JSON
→ evidence extraction
→ structured parsing
→ temporal belief trajectory update
→ evaluation
→ visualization
→ report generation
→ pytest validation

The goal is to check whether temporal evidence can be converted into a belief trajectory over the probability that Event A occurred before Event B.

---

Compared Methods

The benchmark currently compares:

1. Rule-based baseline converter
   A deterministic evidence converter using simple extraction rules.

2. EXAONE scenario-level CoT
   The model receives a full scenario and returns all step judgments at once.

3. EXAONE step-wise evidence v2.1
   The model receives one evidence item at a time and returns a structured judgment.

4. EXAONE order-classification v3
   The model chooses between A_BEFORE_B, B_BEFORE_A, and UNCLEAR.

5. EXAONE cumulative belief v4
   The model receives cumulative evidence and directly estimates the current temporal conclusion.

---

Dataset Contents

Directory	Description
scenarios/	Ten synthetic temporal belief tracking scenarios.
results/	Evaluation outputs, parsed evidence tables, trajectory CSVs, and summary metrics.
figures/	Static visualizations comparing gold trajectories, baseline conversion, and EXAONE/Ollama outputs.
notebooks/	Notebook for inspecting results and reproducing visualizations.
reports/	Markdown experiment report generated from local evaluation outputs.
scripts/	Evaluation, parsing, plotting, and local experiment scripts.
tests/	Application-level usage tests validating expected benchmark files and outputs.

---

Main Result Files

File	Description
results/converter_eval_summary.csv	Baseline converter summary metrics.
results/stepwise_ollama_eval_summary.csv	Stepwise EXAONE/Ollama evaluation summary.
results/stepwise_ollama_scenario_summary.csv	Scenario-level stepwise evaluation results.
results/order_v3_eval_summary.csv	Order-sensitive extraction evaluation summary.
results/cumulative_v4_eval_summary.csv	Cumulative belief prediction evaluation summary.
results/trajectories_gold.csv	Gold belief trajectories.
results/trajectories_auto.csv	Rule-based baseline trajectories.
results/trajectories_stepwise_ollama.csv	Stepwise EXAONE/Ollama belief trajectories.
results/ollama_stepwise_evidence_raw.jsonl	Raw stepwise model outputs.
reports/tbg_cot_experiment_report.md	Generated markdown experiment report.

---

Repository Structure

tbg-cot-bench/
├── scenarios/              # Temporal reasoning benchmark scenarios
├── scripts/                # Experiment, evaluation, visualization, and report scripts
├── results/                # CSV and JSONL experiment outputs
├── figures/                # Generated plots and comparison figures
├── notebooks/              # Result inspection notebook
├── reports/                # Markdown experiment reports
├── tests/                  # Application-level benchmark tests
├── pytest.ini
├── README.md
└── DATASET_CARD.md

---

Notebook

The main inspection notebook is available here:

notebooks/tbg_cot_tracking.ipynb

It can be used to inspect scenario files, load result CSVs, and reproduce trajectory-level visualizations.

---

Requirements

This project was tested locally with:

Ubuntu 22.04
Python virtual environment
Ollama
EXAONE local GGUF model via Ollama
pytest
matplotlib

Install basic Python dependencies:

pip install -r requirements.txt

Ollama should already have a local model registered, for example:

ollama list

Expected model name used in scripts:

exaone-local

---

Running the Benchmark

Activate the virtual environment from the project root:

cd ~/Desktop/CHMLabs/tbg-cot-bench-local-experiment
source ../venv/bin/activate

Run the current best EXAONE step-wise pipeline:

OLLAMA_MODEL=exaone-local bash scripts/run_stepwise_v21_pipeline.sh

Generate the experiment report:

python scripts/generate_experiment_report.py

Run application benchmark tests:

pytest

Expected result:

13 passed

---

Testing

The test suite validates the benchmark as a reproducible application-level experiment.

It checks:

- scenario file structure
- required result files
- step-wise v2.1 performance against the baseline
- trajectory probability bounds
- report generation

Run:

pytest

Current validated result:

13 passed

---

Interpretation

The experiments suggest the following:

1. Scenario-level CoT prompting is unstable for local EXAONE in this setup.
2. Step-wise evidence extraction improves structured-output reliability.
3. Order classification can collapse into UNCLEAR.
4. Cumulative prompting can overload the local model and reduce parse stability.
5. The most reliable current architecture is modular:

evidence extraction
→ structured parsing
→ belief update
→ trajectory evaluation

---

Recommended Next Steps

Planned next steps:

1. Freeze EXAONE step-wise v2.1 as the current best local method.
2. Expand scenarios with harder reversal and noisy convergence cases.
3. Add optional integration tests against the upstream temporal-belief-graph package.
4. Add a lightweight Hugging Face Space only if an interactive demo becomes necessary.
5. Package a stable release after scenario expansion.

---

License

This dataset and its generated benchmark artifacts are released under CC-BY-NC-4.0 unless otherwise specified.

The accompanying code may be relicensed separately depending on release intent.

Recommended code-side options:

Apache-2.0 for research/tooling openness
AGPL-3.0 for stronger defensive sharing requirements

---

Author

Created and maintained by CHML-real.

GitHub:

https://github.com/CHML-real