---

license: apache-2.0

base_model: Qwen/Qwen3.5-0.8B

pipeline_tag: text-generation

language:

- en

tags:

- gguf

- llama-cpp

- lora

- education

- mathematics

- science

---

Muta Tutor Qwen3.5 0.8B Q4_0

This is the scalar-configuration finalist from Muta's ADTC fine-tuning campaign. It starts from

Qwen/Qwen3.5-0.8B, applies BF16 LoRA with rank 16 for 400 steps on multiple-choice math and

science training data, merges the adapter, and exports the result as Q4_0 GGUF.

File

Muta-Tutor-Qwen3.5-0.8B-Q4_0.gguf

SHA-256: 552de22f7ea6f161a458985900e2c961d7578baa1ea9c23018ae27151623ff26

hf download timiiowolabi/Muta-Tutor-Qwen3.5-0.8B-ADTC-GGUF \
  Muta-Tutor-Qwen3.5-0.8B-Q4_0.gguf

Evaluation

The candidate and its zero-learning-rate export control were evaluated with the same GGUF

conversion, quantization, prompt format, and GCP CPU-proxy harness.

| Measure | Control | Fine-tuned |

|---|---:|---:|

| ARC-Easy acc_norm, 500 samples | 55.2% | 70.2% |

| Scalar total score | 72.8896 | 80.3664 |

| Vector total score | 74.8959 | 82.3962 |

Held-out checks improved from 32% to 42% on ARC-Challenge-100, 32% to 33% on

OpenBookQA-100, and 16% to 20% on GSM8K-25.

The repository includes the training manifest, dataset manifest, artifact hashes, and the full

fine-tuning summary. Performance was measured on a cloud CPU proxy; temperature was unavailable,

and peak RSS includes a 45 MiB estimate for the profiler root process.

Status

This remains a competition candidate. It still requires Muta's final embedded tutor template,

live tutoring validation, and profiling on the physical target laptop.