BanglaHalluEval

A hallucination evaluation benchmark for Bengali (Bangla) language models, covering four task types: Question Answering, Summarization, Mathematical Reasoning, and Code-mixed QA. The benchmark evaluates whether model outputs contain hallucinated information, using a suite of judge models ranging from GPT-4.1 mini to locally hosted open-source LLMs.

---

Table of Contents

• Overview
• Tasks
  • 1. Question Answering (QA)
  • 2. Summarization
  • 3. Reasoning
  • 4. Code-mixed QA (Codemix)
• Evaluation Pipeline
• Judge Models
• Evaluation Prompts
• Chain-of-Thought (CoT) Evaluation
• Annotation Quality and Human Validation
• Repository Structure
• Setup

---

Overview

BanglaHalluEval constructs datasets of hallucinated model outputs across four Bengali NLP tasks, then uses judge LLMs to label each sample as hallucinated (yes) or not (no). The benchmark supports two evaluation modes:

• Standard — single-token yes/no judgment
• CoT (Chain-of-Thought) — step-by-step reasoning before a final verdict

Hallucination patterns across tasks include factualness, comprehension, intrinsic, extrinsic, arithmetic_slip, and other error types.

---

Tasks

1. Question Answering (QA)

Bengali reading-comprehension QA. A context passage is provided alongside a question; the model produces a hallucinated answer that is factually wrong or unsupported by the context.

CSV columns:

---

2. Summarization

Bengali abstractive summarization. A source document is summarized; the hallucinated summary either introduces unsupported facts (extrinsic) or contradicts the source (intrinsic).

CSV columns:

---

3. Reasoning

Bengali mathematical word problems drawn from the SOMADHAN dataset. A hallucinated reasoning chain is generated by introducing errors at specific steps of the solution.

CSV columns:

Note: The judge must evaluate the entire reasoning chain, not just the final answer. The TigerLLM and CoT prompts request a JSON response ({"is_hallucinated": "Yes/No"}) for this task instead of a plain token.

---

4. Code-mixed QA (Codemix)

QA over code-switched Bangla–English text (Romanised Bengali mixed with English). The context and question are written in transliterated/mixed script; answers may hallucinate facts not present in the context.

CSV columns:

Example row:

• Context: "1965 saler 31 July England-er Gloucestershire-er Yate shohore jonmogrohon koren J.K. Rowling..."
• Question: J. K. Rowling-er mayer nam ki?
• Ground truth: Anni
• Hallucinated answer: J. K. Rowling-er mayer nam chhilo Margaret Anne.

---

Evaluation Pipeline

Each task follows a two-stage pipeline:

Stage 1 — Hallucination Generation
  Source dataset  →  LLM (GPT-4 / Ollama)  →  hallucinated_answer / hallucinated_chain / hallucinated_summary

Stage 2 — Hallucination Labeling (Judge Models)
  Hallucinated output  →  Judge LLM  →  is_hallucinated: yes / no

Stage 1 scripts are in BanglaHalluEval *Pilot Testing/ and Reasoning/Evaluation Script/.

Stage 2 scripts are in scripts/ and Codemix/.

Both a hallucinated candidates split and a ground-truth (main) split are evaluated, allowing measurement of false-positive rates alongside true hallucination detection rates.

---

Judge Models

All judge models are configured with temperature=0 and a minimal token budget (8–32 tokens) to enforce single-token or structured responses.

---

Evaluation Prompts

QA

You are an evaluator.
Decide whether the provided model answer is hallucinated relative to the question.
Only reply with a single token: yes or no. No explanation, no punctuation, no extra text.
Interpretation: 'yes' means the answer contains information not supported by the question/context
or is factually incorrect (hallucinated).
Provide the answer in English only: yes or no.

Question: {question}
Model answer: {answer}

Answer now:

Summarization

You are an evaluator.
Decide whether the provided model summary is hallucinated relative to the document.
Only reply with a single token: yes or no. No explanation, no punctuation, no extra text.
Interpretation: 'yes' means the summary contains information not supported by the document
or contradicts it (hallucinated).
Provide the answer in English only: yes or no.

Document: {document}
Summary: {summary}

Answer now:

Reasoning

You are an evaluator for Bengali mathematical reasoning tasks.
Decide whether the provided reasoning chain is hallucinated relative to the question.
Only reply with a single token: yes or no. No explanation, no punctuation, no extra text.
Interpretation: 'yes' means the chain or answer contains incorrect steps or unsupported facts.
Provide the answer in English only: yes or no.

Question: {question}
Reasoning Chain: {chain}
Answer: {answer}

Answer now:

TigerLLM uses a JSON variant for this task:

Is this hallucinated_chain hallucinated? Respond ONLY with a JSON object like this:
{"is_hallucinated": "Yes"} or {"is_hallucinated": "No"}

Codemix

You are an evaluator.
Decide whether the provided model answer is hallucinated relative to the context and question.
Only reply with a single token: yes or no. No explanation, no punctuation, no extra text.
Interpretation: 'yes' means the answer contains information not supported by the context
or contradicts it (hallucinated).
Provide the answer in English only: yes or no.

Context: {context}
Question: {question}
Model answer: {answer}

Answer now:

---

Chain-of-Thought (CoT) Evaluation

An alternative evaluation mode (scripts/evaluate_cot_ollama.py) instructs the judge to reason step-by-step before issuing its verdict. This is run across QA, Summarization, and Reasoning tasks using four Ollama models.

QA — CoT

You are an evaluator checking whether a model answer is hallucinated.

Question: {question}
Model Answer: {answer}

Analyze step by step:
Step 1: What factual claims does the answer make?
Step 2: Are these claims supported by or inferable from the question context?
Step 3: Based on steps 1-2, is the answer hallucinated?

Final answer (write only this word on the last line): yes or no
(yes = answer is hallucinated, no = answer is not hallucinated)

Summarization — CoT

You are an evaluator checking whether a summary is hallucinated relative to a document.

Document: {document}
Summary: {summary}

Analyze step by step:
Step 1: List the key claims made in the summary.
Step 2: For each claim, check whether it is directly supported by the document.
Step 3: Based on steps 1-2, decide your final answer.

Final answer (write only this word on the last line): yes or no
(yes = summary is hallucinated, no = summary is not hallucinated)

Reasoning — CoT

You are an expert evaluator for Bengali mathematical reasoning tasks.

Question: {question}
Reasoning Chain: {chain}
Answer: {answer}

Analyze step by step:
Step 1: Is each calculation or logical step in the reasoning chain mathematically correct?
Step 2: Does the final answer follow logically from the reasoning chain?
Step 3: Based on steps 1-2, is this reasoning chain hallucinated (incorrect or fabricated)?

Respond ONLY with a JSON object on the last line:
{"is_hallucinated": "Yes"} or {"is_hallucinated": "No"}

Note: CoT evaluation is not defined for the Codemix task.

CoT models: qwen2.5:32b-instruct, gemma2:27b, mistral-nemo:latest, llama3.1:8b

CoT output files follow the naming pattern *_cot_{model_slug}.csv.

---

Annotation Quality and Human Validation

To validate the benchmark, three annotators independently labelled a blind, stratified sample across all tasks. Annotators saw only the item to be judged (context or document or problem, and the candidate output), were not told whether an item came from the gold or the hallucinated split, and did not confer on individual items during labelling.

Annotator criteria. The complete, task-by-task instructions given to the annotators (definitions, labeling rules, worked examples, and common pitfalls) are in annotation/guidelines/BanHalluEval_Annotation_Guidelines.pdf.

Inter-annotator agreement. Fleiss' kappa across the three annotators, together with two exact agreement rates: the exact agreement and the share of items on which all three annotators agreed unanimously. Figures below reflect a recheck of the annotation process.

*9 items excluded because one annotator skipped them.

Under the Landis and Koch scale, GQA, Code-mixed QA, Summarization, and Reasoning all reach almost perfect agreement (0.911 to 0.926) after the recheck; QA-model outputs (0.568) remains in the moderate range. Per-item agreement tables (each annotator's label, the majority label, unanimity, per-item agreement, and the overall kappa) are produced for every task by the annotation pipeline.

---

Repository Structure

Each task's data lives inside that task's own top-level folder — there is no shared catch-all "Results" or "Datasets" folder anymore (an earlier version of this repo had CSVs scattered across several overlapping locations; they've since been consolidated to one canonical copy per dataset).

BanglaHalluEval/
│
├── Hallucination Generated Answers/       # Raw hallucinated outputs (input to Stage 2), one file per task
│   ├── qa_4000.csv
│   ├── summarization_3000_corrected.csv
│   ├── reasoning_1000.csv
│   └── codemix_4000.csv
│
├── BanglaHalluEval Datasets/              # Final curated 1000-sample QA ground-truth dataset
│   └── banglahallueval_qa_1000.csv
│
├── QA/
│   └── Results/                           # QA evaluation/labeling output CSVs
│
├── Summarization/
│   ├── 1000 Selected Samples/             # Final curated 1000-sample summarization ground truth
│   ├── 3000 Candidate Answers/
│   └── Results/                           # Summarization evaluation/labeling output CSVs
│
├── Reasoning/
│   ├── SOMADHAN Source Dataset/           # Raw SOMADHAN.csv -> length-sorted -> extreme-sample exploration
│   │   ├── SOMADHAN.csv
│   │   ├── SOMADHAN_SORTED.csv
│   │   ├── sort_dataset.py
│   │   └── create_extreme_samples.py
│   ├── 1000 Selected Samples/             # Final curated 1000-sample reasoning ground truth
│   ├── Evaluation Script/
│   └── Results/                           # Reasoning evaluation/labeling output CSVs
│
├── Codemix/
│   ├── Main dataset/                      # Final curated 1000-sample codemix ground truth
│   ├── Results/                           # Codemix evaluation/labeling output CSVs
│   └── label_codemix_*_ollama.py          # Per-model Codemix labeling scripts
│
├── Sample Selection for QA/                # QA candidate-pool + selection stage (pre-dates the final 1000)
├── Sample Selection for Summ/              # Summarization candidate-pool + selection stage
│
├── scripts/                               # All evaluation and labeling scripts
│   ├── label_hallucinations_ollama.py     # QA — Ollama judge
│   ├── label_hallucinations_gpt.py        # QA — GPT judge
│   ├── label_correct_answers_gpt_4_1_mini.py
│   ├── label_summarization_ollama.py      # Summarization — Ollama judge
│   ├── label_summarization_gpt_4_1_mini.py
│   ├── label_summarization_gemma2.py
│   ├── label_summarization_deepseek.py
│   ├── label_reasoning_gpt_4_1_mini.py    # Reasoning — GPT judge
│   ├── label_reasoning_main_1000_gpt_4_1_mini.py
│   ├── label_codemix_hallucinations_gpt_4_1_mini.py  # Codemix — GPT judge
│   ├── label_codemix_main_gpt_4_1_mini.py
│   ├── evaluate_tigerllm.py               # TigerLLM — all tasks (QA/Summ/Reasoning)
│   ├── evaluate_cot_ollama.py             # CoT evaluation — QA/Summ/Reasoning
│   ├── evaluate_deepseek_ollama.py        # DeepSeek — QA/Summ
│   ├── evaluate_codemix_deepseek_candidates.py
│   ├── evaluate_codemix_deepseek_main.py
│   ├── evaluate_reasoning_deepseek_candidates.py
│   ├── evaluate_reasoning_deepseek_main.py
│   └── legacy/                            # Superseded single-machine TigerLLM scripts, kept for history
│
├── Evaluation/                            # GPT-4.1-mini/LLaMA baseline results (QA + Reasoning)
├── T Sampled Evaluations/                  # 10%-sample baseline+CoT runs (BanglaLLaMA/TituLLM comparison)
├── full_ecot_run/, pilot_50_samples/       # E-CoT hallucination-mitigation pipeline (see HALLUCINATION_MITIGATION.md)
├── annotation/guidelines/                  # Human-annotation instructions
├── public/halueval.pdf                    # Reference paper (Li et al., HaluEval)
├── requirements.txt
└── README.md

---

Setup

1. Install dependencies

pip install -r requirements.txt
pip install openai python-dotenv pandas transformers torch accelerate

2. Configure API keys (for GPT-based scripts)

Create a .env file in the project root:

OPENAI_API_KEY=sk-...

3. Start Ollama (for Ollama-based scripts)

ollama serve
ollama pull qwen2.5:32b-instruct

For a remote RunPod instance, pass --ollama-url https://<pod-id>-11434.proxy.runpod.net.

4. Run an evaluation

# QA — GPT-4.1 mini
python scripts/label_hallucinations_gpt.py \
  --input "Hallucination Generated Answers/qa_4000.csv" \
  --output "QA/Results/qa_4000_gpt4mini.csv"

# Summarization — Qwen 32B via Ollama
python scripts/label_summarization_ollama.py \
  --input "Hallucination Generated Answers/summarization_3000_corrected.csv" \
  --output "Summarization/Results/summ_3000_qwen32b.csv" \
  --model qwen2.5:32b-instruct

# Reasoning — GPT-4.1 mini
python scripts/label_reasoning_gpt_4_1_mini.py \
  --input "Hallucination Generated Answers/reasoning_1000.csv" \
  --output "Reasoning/Results/reasoning_1000_gpt4mini.csv"

# Codemix — GPT-4.1 mini
python scripts/label_codemix_hallucinations_gpt_4_1_mini.py \
  --input "Hallucination Generated Answers/codemix_4000.csv" \
  --output "Codemix/Results/codemix_4000_gpt4mini.csv"

# TigerLLM — all tasks (requires local GPU)
python scripts/evaluate_tigerllm.py --task all

# CoT — all tasks, all models
python scripts/evaluate_cot_ollama.py --task all

All scripts support --resume (skip already-labeled rows), --start / --end (slice), and --force-binary (retry until yes/no is obtained).