evaluating-llms-harness

lm-evaluation-harness - LLM Benchmarking

Quick start

lm-evaluation-harness evaluates LLMs across 60+ academic benchmarks using standardized prompts and metrics.

Installation:

pip install lm-eval

Evaluate any HuggingFace model:

lm_eval --model hf \
  --model_args pretrained=meta-llama/Llama-2-7b-hf \
  --tasks mmlu,gsm8k,hellaswag \
  --device cuda:0 \
  --batch_size 8

View available tasks:

lm_eval --tasks list

Common workflows

Workflow 1: Standard benchmark evaluation

Evaluate model on core benchmarks (MMLU, GSM8K, HumanEval).

Copy this checklist:

Benchmark Evaluation:
- [ ] Step 1: Choose benchmark suite
- [ ] Step 2: Configure model
- [ ] Step 3: Run evaluation
- [ ] Step 4: Analyze results

Step 1: Choose benchmark suite

Core reasoning benchmarks:

• MMLU (Massive Multitask Language Understanding) - 57 subjects, multiple choice
• GSM8K - Grade school math word problems
• HellaSwag - Common sense reasoning
• TruthfulQA - Truthfulness and factuality
• ARC (AI2 Reasoning Challenge) - Science questions

Code benchmarks:

• HumanEval - Python code generation (164 problems)
• MBPP (Mostly Basic Python Problems) - Python coding

Standard suite (recommended for model releases):

--tasks mmlu,gsm8k,hellaswag,truthfulqa,arc_challenge

Step 2: Configure model

HuggingFace model:

lm_eval --model hf \
  --model_args pretrained=meta-llama/Llama-2-7b-hf,dtype=bfloat16 \
  --tasks mmlu \
  --device cuda:0 \
  --batch_size auto  # Auto-detect optimal batch size

Quantized model (4-bit/8-bit):

lm_eval --model hf \
  --model_args pretrained=meta-llama/Llama-2-7b-hf,load_in_4bit=True \
  --tasks mmlu \
  --device cuda:0

Custom checkpoint:

lm_eval --model hf \
  --model_args pretrained=/path/to/my-model,tokenizer=/path/to/tokenizer \
  --tasks mmlu \
  --device cuda:0

Step 3: Run evaluation

# Full MMLU evaluation (57 subjects)
lm_eval --model hf \
  --model_args pretrained=meta-llama/Llama-2-7b-hf \
  --tasks mmlu \
  --num_fewshot 5 \  # 5-shot evaluation (standard)
  --batch_size 8 \
  --output_path results/ \
  --log_samples  # Save individual predictions

# Multiple benchmarks at once
lm_eval --model hf \
  --model_args pretrained=meta-llama/Llama-2-7b-hf \
  --tasks mmlu,gsm8k,hellaswag,truthfulqa,arc_challenge \
  --num_fewshot 5 \
  --batch_size 8 \
  --output_path results/llama2-7b-eval.json

Step 4: Analyze results

Results saved to results/llama2-7b-eval.json:

{
  "results": {
    "mmlu": {
      "acc": 0.459,
      "acc_stderr": 0.004
    },
    "gsm8k": {
      "exact_match": 0.142,
      "exact_match_stderr": 0.006
    },
    "hellaswag": {
      "acc_norm": 0.765,
      "acc_norm_stderr": 0.004
    }
  },
  "config": {
    "model": "hf",
    "model_args": "pretrained=meta-llama/Llama-2-7b-hf",
    "num_fewshot": 5
  }
}

Workflow 2: Track training progress

Evaluate checkpoints during training.

Training Progress Tracking:
- [ ] Step 1: Set up periodic evaluation
- [ ] Step 2: Choose quick benchmarks
- [ ] Step 3: Automate evaluation
- [ ] Step 4: Plot learning curves

Step 1: Set up periodic evaluation

Evaluate every N training steps:

#!/bin/bash
# eval_checkpoint.sh

CHECKPOINT_DIR=$1
STEP=$2

lm_eval --model hf \
  --model_args pretrained=$CHECKPOINT_DIR/checkpoint-$STEP \
  --tasks gsm8k,hellaswag \
  --num_fewshot 0 \  # 0-shot for speed
  --batch_size 16 \
  --output_path results/step-$STEP.json

Step 2: Choose quick benchmarks

Fast benchmarks for frequent evaluation:

• HellaSwag: ~10 minutes on 1 GPU
• GSM8K: ~5 minutes
• PIQA: ~2 minutes

Avoid for frequent eval (too slow):

• MMLU: ~2 hours (57 subjects)
• HumanEval: Requires code execution

Step 3: Automate evaluation

Integrate with training script:

# In training loop
if step % eval_interval == 0:
    model.save_pretrained(f"checkpoints/step-{step}")

    # Run evaluation
    os.system(f"./eval_checkpoint.sh checkpoints step-{step}")

Or use PyTorch Lightning callbacks:

from pytorch_lightning import Callback

class EvalHarnessCallback(Callback):
    def on_validation_epoch_end(self, trainer, pl_module):
        step = trainer.global_step
        checkpoint_path = f"checkpoints/step-{step}"

        # Save checkpoint
        trainer.save_checkpoint(checkpoint_path)

        # Run lm-eval
        os.system(f"lm_eval --model hf --model_args pretrained={checkpoint_path} ...")

Step 4: Plot learning curves

import json
import matplotlib.pyplot as plt

# Load all results
steps = []
mmlu_scores = []

for file in sorted(glob.glob("results/step-*.json")):
    with open(file) as f:
        data = json.load(f)
        step = int(file.split("-")[1].split(".")[0])
        steps.append(step)
        mmlu_scores.append(data["results"]["mmlu"]["acc"])

# Plot
plt.plot(steps, mmlu_scores)
plt.xlabel("Training Step")
plt.ylabel("MMLU Accuracy")
plt.title("Training Progress")
plt.savefig("training_curve.png")

Workflow 3: Compare multiple models

Benchmark suite for model comparison.

Model Comparison:
- [ ] Step 1: Define model list
- [ ] Step 2: Run evaluations
- [ ] Step 3: Generate comparison table

Step 1: Define model list

# models.txt
meta-llama/Llama-2-7b-hf
meta-llama/Llama-2-13b-hf
mistralai/Mistral-7B-v0.1
microsoft/phi-2

Step 2: Run evaluations

#!/bin/bash
# eval_all_models.sh

TASKS="mmlu,gsm8k,hellaswag,truthfulqa"

while read model; do
    echo "Evaluating $model"

    # Extract model name for output file
    model_name=$(echo $model | sed 's/\//-/g')

    lm_eval --model hf \
      --model_args pretrained=$model,dtype=bfloat16 \
      --tasks $TASKS \
      --num_fewshot 5 \
      --batch_size auto \
      --output_path results/$model_name.json

done < models.txt

Step 3: Generate comparison table

import json
import pandas as pd

models = [
    "meta-llama-Llama-2-7b-hf",
    "meta-llama-Llama-2-13b-hf",
    "mistralai-Mistral-7B-v0.1",
    "microsoft-phi-2"
]

tasks = ["mmlu", "gsm8k", "hellaswag", "truthfulqa"]

results = []
for model in models:
    with open(f"results/{model}.json") as f:
        data = json.load(f)
        row = {"Model": model.replace("-", "/")}
        for task in tasks:
            # Get primary metric for each task
            metrics = data["results"][task]
            if "acc" in metrics:
                row[task.upper()] = f"{metrics['acc']:.3f}"
            elif "exact_match" in metrics:
                row[task.upper()] = f"{metrics['exact_match']:.3f}"
        results.append(row)

df = pd.DataFrame(results)
print(df.to_markdown(index=False))

Output:

| Model                  | MMLU  | GSM8K | HELLASWAG | TRUTHFULQA |
|------------------------|-------|-------|-----------|------------|
| meta-llama/Llama-2-7b  | 0.459 | 0.142 | 0.765     | 0.391      |
| meta-llama/Llama-2-13b | 0.549 | 0.287 | 0.801     | 0.430      |
| mistralai/Mistral-7B   | 0.626 | 0.395 | 0.812     | 0.428      |
| microsoft/phi-2        | 0.560 | 0.613 | 0.682     | 0.447      |

Workflow 4: Evaluate with vLLM (faster inference)

Use vLLM backend for 5-10x faster evaluation.

vLLM Evaluation:
- [ ] Step 1: Install vLLM
- [ ] Step 2: Configure vLLM backend
- [ ] Step 3: Run evaluation

Step 1: Install vLLM

pip install vllm

Step 2: Configure vLLM backend

lm_eval --model vllm \
  --model_args pretrained=meta-llama/Llama-2-7b-hf,tensor_parallel_size=1,dtype=auto,gpu_memory_utilization=0.8 \
  --tasks mmlu \
  --batch_size auto

Step 3: Run evaluation

vLLM is 5-10× faster than standard HuggingFace:

# Standard HF: ~2 hours for MMLU on 7B model
lm_eval --model hf \
  --model_args pretrained=meta-llama/Llama-2-7b-hf \
  --tasks mmlu \
  --batch_size 8

# vLLM: ~15-20 minutes for MMLU on 7B model
lm_eval --model vllm \
  --model_args pretrained=meta-llama/Llama-2-7b-hf,tensor_parallel_size=2 \
  --tasks mmlu \
  --batch_size auto

When to use vs alternatives

Use lm-evaluation-harness when:

• Benchmarking models for academic papers
• Comparing model quality across standard tasks
• Tracking training progress
• Reporting standardized metrics (everyone uses same prompts)
• Need reproducible evaluation

Use alternatives instead:

• HELM (Stanford): Broader evaluation (fairness, efficiency, calibration)
• AlpacaEval: Instruction-following evaluation with LLM judges
• MT-Bench: Conversational multi-turn evaluation
• Custom scripts: Domain-specific evaluation

Common issues

Details, code examples and parameter tables: references/common-issues.md. Read it when this step applies.

Advanced topics

Benchmark descriptions: See references/benchmark-guide.md for detailed description of all 60+ tasks, what they measure, and interpretation.

Custom tasks: See references/custom-tasks.md for creating domain-specific evaluation tasks.

API evaluation: See references/api-evaluation.md for evaluating OpenAI, Anthropic, and other API models.

Multi-GPU strategies: See references/distributed-eval.md for data parallel and tensor parallel evaluation.

Hardware requirements

Details, code examples and parameter tables: references/hardware-requirements.md. Read it when this step applies.

Resources

• GitHub: https://github.com/EleutherAI/lm-evaluation-harness
• Docs: https://github.com/EleutherAI/lm-evaluation-harness/tree/main/docs
• Task library: 60+ tasks including MMLU, GSM8K, HumanEval, TruthfulQA, HellaSwag, ARC, WinoGrande, etc.
• Leaderboard: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard (uses this harness)

Agent operating procedure

1. Check the environment. Confirm the model, checkpoint, benchmark version and evaluation harness version.
2. Pin down the inputs. Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.
3. Run a small version first. Evaluate on a small subset to verify prompts, parsing and scoring.
4. Execute the full task using the instructions and references above.
5. Validate the result. Use the benchmark's official splits and metrics; check for contamination; report variance across seeds or prompts.
6. Report. State what was run (versions, commands, parameters), what was checked, and what is still uncertain.

If this happens	Do this
Scores differ from published numbers	Check prompt format, few-shot settings and harness version before concluding.
A function, flag or endpoint in these instructions is missing in the installed version	Check the installed version's own documentation (help(), --help, official docs), adapt, and tell the user. Never invent an API.
A required input, identifier or parameter is ambiguous	Ask the user, or state the assumption explicitly before running.

Integrity rules

• Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.
• Report evaluation settings in full; never cherry-pick prompts, seeds or subsets.
• Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.
• Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.

Related skills

• evaluating-code-models: Evaluates code generation models across HumanEval, MBPP, MultiPL-E, and 15+ benchmarks with pass@k metrics.
• nemo-evaluator-sdk: Evaluates LLMs across 100+ benchmarks from 18+ harnesses (MMLU, HumanEval, GSM8K, safety, VLM) with multi-backend execution.
• hqq-quantization: Half-Quadratic Quantization for LLMs without calibration data.