---

license: apache-2.0

base_model: Qwen/Qwen2.5-1.5B-Instruct

pipeline_tag: text-generation

language:

- en

tags:

- gguf

- llama-cpp

- lora

- education

- mathematics

- science

---

Muta Tutor Qwen2.5 1.5B Q4_K_M

Final submission: Muta-Tutor-Qwen2.5-1.5B-Q4_K_M-vocab32k.gguf

Muta's final ADTC submission is the vocabulary-pruned version of the fine-tuned Q4_K_M below.

The tokenizer and embedding/output rows were cut from 151,936 to 32,000 tokens (merges

151,387 → 31,722). The kept set covers 100% of the 4.05M tokens in the training corpus. The

weights and quantization are otherwise unchanged. The pruning manifest, with the source and

artifact hashes, the corpus hashes and the tokenizer revision, is

Muta-Tutor-Qwen2.5-1.5B-Q4_K_M-vocab32k.vocab-prune-manifest.json.

• Size: 830,115,168 bytes (source: 986,048,128)
• SHA-256: 3038036a93039e80d282581d451abfdf998cd2929d2d26720a2cc711bff33a94
• Source SHA-256: a750d00d458c6ab38925364ea1413db00648449180941e47025736d09922e1eb

hf download timiiowolabi/Muta-Tutor-Qwen2.5-1.5B-ADTC-GGUF \
  Muta-Tutor-Qwen2.5-1.5B-Q4_K_M-vocab32k.gguf

Source model

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

Qwen/Qwen2.5-1.5B-Instruct, applies BF16 LoRA with rank 16 for 500 steps on a licence-clean

multiple-choice math and science mixture, merges the adapter, and exports the result as Q4_K_M

GGUF.

File

Muta-Tutor-Qwen2.5-1.5B-Q4_K_M.gguf

SHA-256: a750d00d458c6ab38925364ea1413db00648449180941e47025736d09922e1eb

hf download timiiowolabi/Muta-Tutor-Qwen2.5-1.5B-ADTC-GGUF \
  Muta-Tutor-Qwen2.5-1.5B-Q4_K_M.gguf

Evaluation

The candidate and untuned 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 | 74.4% | 77.8% |

| Scalar total score | 65.3277 | 67.0475 |

| Vector total score | 82.4386 | 84.1387 |

The repository includes the training manifest, licence-clean 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. The secondary

held-out battery is pending because its first attempt overlapped unrelated CPU work and was

discarded.

Status

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

live tutoring validation, the secondary held-out battery, and profiling on the physical target

laptop.