license: mit
language:
- en
tags:
- quantization
- evaluation
- small-language-models
- benchmark-results
- lm-evaluation-harness
pretty_name: Capability-Specific Degradation of Quantized Small Language Models
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: processed/results_long.csv
Capability-Specific Degradation of Quantized Small Language Models — raw evaluation outputs
Raw and aggregated evaluation results for the study Capability-Specific Degradation Patterns in Quantized Small Language Models: seven open instruction-tuned small LLMs (1–4B parameters, five architecture families) evaluated at FP16 and 4-bit (bitsandbytes NF4) across six capabilities, for 84 controlled model × precision × benchmark evaluations.
This repository contains evaluation outputs only — no model weights and no benchmark data are redistributed here.
Contents
| Path | What it is |
|---|---|
raw/<model>/<precision>__<benchmark>.json |
One unmodified lm-evaluation-harness result record per evaluation, e.g. raw/gemma-2-2b-it/fp16__gsm8k.json. Each holds the task metrics, standard errors, sample counts, and the full harness config (model args, precision, few-shot setting, batch size). |
processed/results_long.csv |
Tidy long-format table: one row per model × precision × benchmark × metric. |
processed/results_table.csv |
Wide table of headline scores per model × precision. |
The JSON records are the primary artifact; the CSVs are derived from them by
scripts/aggregate_results.py in the code repository.
Models
Qwen/Qwen2.5-3B-Instruct, Qwen/Qwen2.5-1.5B-Instruct, meta-llama/Llama-3.2-3B-Instruct,
meta-llama/Llama-3.2-1B-Instruct, google/gemma-2-2b-it, microsoft/Phi-3.5-mini-instruct,
HuggingFaceTB/SmolLM2-1.7B-Instruct.
Capabilities and benchmarks
| Capability | Benchmark | Metric |
|---|---|---|
| Factual knowledge | ARC-Challenge | acc_norm |
| Commonsense reasoning | HellaSwag | acc_norm |
| Mathematical reasoning | GSM8K | exact_match |
| Multilingual math | MGSM | exact_match |
| Code generation | HumanEval | pass@1 |
| Instruction following | IFEval | prompt-level strict acc |
Generative tasks were run on vLLM and multiple-choice log-likelihood tasks on the Hugging Face backend; the same backend is used for a given capability across all models and precisions, so score differences reflect quantization rather than the harness.
Headline result
Relative degradation FP16 → 4-bit (%, higher is worse):
| Model | GSM8K | ARC | HellaSwag | HumanEval | IFEval | MGSM |
|---|---|---|---|---|---|---|
| Qwen2.5-3B | −2.6 | 3.4 | 2.6 | 9.2 | 1.9 | 19.0 |
| Qwen2.5-1.5B | 10.6 | 9.3 | 2.6 | 22.6 | 11.8 | 22.9 |
| Llama-3.2-3B | 4.0 | 3.1 | 1.8 | −6.1 | −0.5 | 19.1 |
| Llama-3.2-1B | 16.7 | 2.7 | 5.1 | 12.3 | 4.5 | — |
| Gemma-2-2B | 4.1 | 2.4 | 1.8 | −2.6 | 1.0 | 12.2 |
| Phi-3.5-mini | 4.3 | −1.7 | 1.7 | 21.0 | 1.1 | 19.0 |
| SmolLM2-1.7B | 18.1 | 2.7 | 2.4 | 57.2† | −1.2 | — |
† chat-template prompting protocol. "—" = FP16 baseline too low for a stable ratio.
Degradation is strongly capability-specific — multilingual math and code are most fragile, commonsense is almost untouched — but the ordering is only weakly consistent across architectures (mean pairwise Kendall's τ = 0.29).
Usage
from huggingface_hub import snapshot_download
snapshot_download("Emil-7/llm-quant-degradation", repo_type="dataset",
local_dir="results_download", allow_patterns=["raw/**", "processed/**"])
The dataset viewer preview may fail because the per-model JSON records have heterogeneous schemas (each record carries only the tasks that run produced). Download the files directly as above rather than relying on the viewer.
Known limitations
Single evaluation runs (no repeated seeds); one quantization method (NF4) at one bit-width; one benchmark per capability. Standard errors reported by the harness are included in every JSON record.
Citation
@misc{rahimov2026quantdegradation,
title = {Capability-Specific Degradation Patterns in Quantized Small Language Models},
author = {Rahimov, Emil},
year = {2026},
note = {Evaluation results dataset},
url = {https://huggingface.co/datasets/Emil-7/llm-quant-degradation}
}
License
MIT, for the evaluation outputs in this repository. The evaluated models and the benchmark datasets remain under their own respective licenses.