---
license: apache-2.0
pretty_name: AWQ Quant Quality Evals
tags:
  - evaluation
  - quantization
  - awq
  - openllm
  - mteb
  - qwen
  - vllm
task_categories:
  - text-generation
  - feature-extraction
size_categories:
  - n<1K
configs:
  - config_name: summary
    data_files: summary.jsonl
---

# AWQ Quant Quality Evals

Side-by-side **AWQ W4A16 vs BF16** quality measurements for:

| Quant | Base | Suite |
|---|---|---|
| [LostGentoo/Qwen3.5-4B-AWQ](https://huggingface.co/LostGentoo/Qwen3.5-4B-AWQ) | Qwen/Qwen3.5-4B | OpenLLM-lite |
| [LostGentoo/Qwen3-Embedding-8B-AWQ](https://huggingface.co/LostGentoo/Qwen3-Embedding-8B-AWQ) | Qwen/Qwen3-Embedding-8B | MTEB-lite |

Hardware: NVIDIA RTX 5060 Ti (sm_120, Blackwell).

## Files

| File | Contents |
|---|---|
| `quant_quality_evals.json` | Full combined report |
| `qwen35_4b_awq_vs_bf16.json` | LLM OpenLLM-lite only |
| `qwen3_embedding_8b_awq_vs_bf16.json` | Embedding MTEB-lite only |
| `summary.jsonl` | Flat per-task rows (easy to load) |

## Headline results

### Qwen3.5-4B (OpenLLM-lite, n=200 MC / task)

| Task | BF16 | AWQ | Recovery |
|---|---:|---:|---:|
| MMLU | 0.710 | 0.700 | 98.6% |
| ARC-Challenge | 0.545 | 0.555 | 101.8% |
| HellaSwag | 0.705 | 0.670 | 95.0% |
| Winogrande | 0.685 | 0.675 | 98.5% |
| TruthfulQA MC1 | 0.295 | 0.285 | 96.6% |
| WikiText-2 PPL | 16.77 | 18.54 | +10.6% |

Mean MC recovery: **98.1%**.

GSM8K in this dump is **not reliable** (short gen budget + weak numeric extraction); treat as a harness caveat, not a quant cliff.

### Qwen3-Embedding-8B (MTEB-lite)

| Task | BF16 | AWQ | Recovery |
|---|---:|---:|---:|
| STSBenchmark Spearman | 0.884 | 0.881 | 99.6% |
| SciFact nDCG@10 | 0.787 | 0.777 | 98.8% |
| STS score Pearson AWQ vs BF16 | - | - | 0.996 |

## Methodology

- **LLM MC tasks**: greedy average-logprob scoring of answer continuations (ARC, HellaSwag, Winogrande, MMLU, TruthfulQA MC1), n=200 shuffled with seed 42.
- **WikiText-2 PPL**: manual NLL from logits, 32 non-empty docs, max length 512.
- **Embeddings**: last-token pooling + L2 normalize; retrieval queries use Qwen instruct prefix; SciFact nDCG@10 over test qrels.
- Harness: custom script on top of cached HF datasets (lm-eval / mteb package install was unavailable in the run environment).

## Load

```python
from datasets import load_dataset
import json
from huggingface_hub import hf_hub_download

rows = load_dataset("LostGentoo/awq-quant-evals", split="train")
print(rows[0])

path = hf_hub_download("LostGentoo/awq-quant-evals", "quant_quality_evals.json", repo_type="dataset")
report = json.load(open(path))
print(report["qwen35_4b"]["recovery_pct"])
```

## License

Apache-2.0. Eval numbers are measurements of third-party base models under their own licenses.