EVAL.md

# IFEval-Hi Evaluation Framework

## Overview

IFEval-Hi is a Hindi language adaptation of the IFEval (Instruction Following Evaluation) benchmark, designed to evaluate the instruction-following capabilities of Large Language Models (LLMs) in Hindi. This implementation maintains the core evaluation methodology of the original English IFEval while incorporating language-specific modifications to ensure accurate and fair assessment of Hindi language models.

## Getting Started

You have two options to use this evaluation framework:

1. **Option 1: Use the Ready-to-Use Fork** (Recommended)
   - Fork or clone the repository directly from: https://github.com/anushaknvidia/lm-evaluation-harness
   - This fork already includes all the Hindi-specific configurations and modifications
   - Skip to [Step 3: Run Evaluation](#step-3-run-evaluation)

2. **Option 2: Manual Setup**
   - Follow the step-by-step instructions below to set up IFEval-Hi from scratch
   - This is useful if you want to customize or understand the implementation details

## Setup and Usage

### Step 1: Create Task Configuration

1. Navigate to the lm-evaluation-harness tasks directory:
   ```
   https://github.com/EleutherAI/lm-evaluation-harness/tree/main/lm_eval/tasks/ifeval
   ```

2. Create a copy of the English IFEval directory and rename it to ifevalhi 

3. Rename the task file in the copied folder to `ifevalhi.yaml` for Hindi-specific configuration

### Step 2: Configure Parameters

Update the `ifevalhi.yaml` configuration file with the following Hindi-specific parameters:

```yaml
# Dataset Configuration
dataset_path: nvidia/IFEval-Hi

# Generation Parameters
max_gen_toks: 4096  # Increased from 1280 to accommodate Hindi morphology

# Additional Hindi-specific settings
# (Include language-specific preprocessing and normalization settings as needed)
```

**Key Configuration Changes:**
- **`dataset_path`**: Changed from `google/IFEval` to `nvidia/IFEval-Hi`
- **`max_gen_toks`**: Increased to 4096 tokens to handle Hindi's linguistic complexity

### Step 3: Run Evaluation

Execute the evaluation using the lm-eval-harness framework with the Hindi task configuration:

```bash
# Basic evaluation command add other arguments as per lm-eval-harness repo
lm-eval --model hf \
        --model_args pretrained=<model_name_or_path> \
        --tasks ifevalhi \
        --batch_size auto \
        --output_path ./results/
```

### Expected Output

The evaluation will generate results including:
- **prompt_level_strict_acc**: Primary accuracy metric
- **normalised_acc**: Normalized accuracy with text preprocessing

## Key Differences from English IFEval

### 1. Configuration Parameters

#### Maximum Generation Token Limit
- **English IFEval**: 1,280 tokens
- **IFEval-Hindi**: 4,096 tokens

The increased token limit accommodates the morphological and syntactic properties of Hindi text, which often requires more tokens to express equivalent content compared to English.

### 2. Language-Specific Processing

#### Tokenization and Segmentation
- **English Implementation**: Uses standard tokenizer for sentence and word segmentation
- **IFEval-Hi**: Incorporates Hindi-specific punctuation handling, including:
  - Sentence delimitation using the vertical bar (`|`) character
  - Custom punctuation rules tailored to Hindi text structure

### 3. Constrained Response Categories

IFEval-Hi expands the constrained response category with Hindi-specific response patterns:

```
- मेरा जवाब हाँ है (My answer is yes)
- मेरा जवाब नहीं है (My answer is no)
- मेरा जवाब शायद है (My answer is maybe)
- हाँ (Yes)
- नहीं (No)
- शायद (Maybe)
```

These additions ensure fair evaluation for Hindi responses and align with natural Hindi language usage patterns.

### 4. Text Normalization

IFEval-Hindi implements comprehensive normalization procedures for model-generated Hindi text and evaluation parameters:

#### Character Normalization
- **Consonant Unification**: Characters like क़ and क are unified to maintain consistency
- **Diacritic Removal**: Diacritical marks such as "ँ" (chandrabindu) are stripped
- **Symbol Cleanup**: Redundant symbols and spacing irregularities are removed
- **Orthographic Standardization**: Variations in Hindi script representation are normalized

These normalization steps ensure consistent processing across input prompts and model-generated outputs, reducing evaluation bias from orthographic variations.

### 5. Validation Logic Updates

#### Letter Frequency Checker
- **English IFEval**: Includes English alphabet-only validation logic
- **IFEval-Hi**: English alphabet validation has been deprecated and removed from `instructions.py` to align with Hindi-specific evaluation requirements

This modification ensures that character-level constraints are appropriately evaluated for the Devanagari script used in Hindi.


IFEval-Hi follows the same execution pipeline as the English variant within the lm-eval-harness repository

```
Pipeline Structure:
1. Load dataset from nvidia/IFEval-Hi
2. Generate model responses with Hindi-specific configurations
3. Apply Hindi text normalization
4. Evaluate instruction-following accuracy
5. Report metrics
```

Both implementations utilize the same core Python utility modules, ensuring consistency in evaluation methodology while supporting language-specific adaptations.
Please find the fork to the evaluation repo with the above changes here https://github.com/anushaknvidia/lm-evaluation-harness