---
tags:
- causal-lm
- qwen
- fine-tuned
- quantized
- mnlp
---

# Qwen3-0.6B Full-Precision + W8A8 Quantized MCQA Model

**Repository:** [Kikinoking/MNLP_M2_quantized_model](https://huggingface.co/Kikinoking/MNLP_M2_quantized_model)

This is a fine-tuned Qwen-3-0.6B causal-LM, trained on a concatenation of multiple MCQA datasets and then quantized to 8-bit weights and activations using the compressed-tensors format. It is designed for multiple-choice QA tasks, evaluated with the LightEval EPFL MNLP suite.

---

## Model Details

- **Base architecture:** Qwen-3 (0.6B parameters)
- **Pretrained checkpoint:** `Qwen/Qwen3-0.6B-Base`
- **Fine-tuning data sources:**
  - ScienceQA
  - QASC
  - OpenBookQA
  - MathQA
  - CommonsenseQA
  - MCQA prompts generated via ChatGPT (labeled `M1_chatgpt`)
- **Dataset split:** 95% train / 5% validation
- **Tokenization:**
  - `AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B-Base")`
  - Left padding, EOS token as pad_token
  - Sequence length capped at 2048 tokens

---

## Quantization

- **Method:** `compressed-tensors` / `naive-quantized`
- **Precision:** 8-bit weights + 8-bit activations
- **Layers kept in FP32:** Language modeling head
- **Checkpoint:** Compatible with CPU and GPU inference

---

## Evaluation

Tested using LightEval EPFL MNLP on the MCQA task:

```bash
lighteval accelerate   --eval-mode lighteval   --save-details   --override-batch-size 8   --custom-tasks community_tasks/mnlp_mcqa_evals.py   --output-dir out/lighteval_quant   model_configs/quantized_model.yaml   "community|mnlp_mcqa_evals|0|0"

Results:

    Accuracy: 0.30 ± 0.15

    Normalized Accuracy: 0.30 ± 0.15

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained(
    "Kikinoking/MNLP_M2_quantized_model", trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
    "Kikinoking/MNLP_M2_quantized_model",
    trust_remote_code=True,
    device_map="auto",
)

License

    Being a 0.6B-parameter model, it may struggle with very long or ambiguous queries.

    Quantization can introduce a slight drop in accuracy (~5–10%).

    License: CC BY-NC 4.0 (inherits from the base Qwen-3 license)