ZAYA1-8B-Coder-GGUF

ZAYA1-8B Coder GGUF

GGUF quantizations of josephmayo/ZAYA1-8B-Coder, the merged Coder model from Zyphra/ZAYA1-8B plus josephmayo/ZAYA1-8B-Coder-LoRA.

Evaluation Gate

The LoRA was evaluated against the base model on 50 Python code-generation prompts with a 0-10 heuristic score:

• Base average: 2.36 / 10
• LoRA average: 4.76 / 10
• Absolute score delta: +2.40 / 10
• Full-scale lift: 24.00%
• Relative lift over base average: 101.69%
• Improved prompts: 39 / 50
• Merge threshold: 20.00%
• Merge decision: true

Full-scale lift is the required notebook metric:

((lora_avg - base_avg) / 10) * 100
((4.76 - 2.36) / 10) * 100 = 24.00%

Architecture And Conversion Notes

ZAYA uses general.architecture = zaya in GGUF. Mainline llama.cpp did not recognize that architecture during quantization, so these files were generated with the experimental ZAYA llama.cpp branch that includes zaya.cpp model support.

The conversion path was:

1. Evaluate base vs LoRA on 50 Python prompts.
2. Merge the adapter after the 24.00% full-scale lift passed the 20.00% threshold.
3. Save the merged model to Hugging Face safetensors.
4. Convert merged safetensors to FP16 GGUF with ZAYA-aware convert_hf_to_gguf.py.
5. Quantize the FP16 GGUF with the ZAYA-aware llama-quantize.

Kaggle completed the eval and merged-model upload, but Kaggle disk was not large enough to hold the merged shards, FP16 GGUF, and quant outputs at the same time. GGUF quantization was completed locally with the same ZAYA llama.cpp branch and then pushed to this repo.

Files

• zaya1-8b-coder-q4_k_m.gguf
• zaya1-8b-coder-q6_k.gguf
• zaya1-8b-coder-q8_0.gguf
• zaya1_8b_coder_gguf_summary.json

File Sizes

• Q4_K_M: 5,567,581,024 bytes
• Q6_K: 7,353,195,104 bytes
• Q8_0: 9,485,673,760 bytes

<!-- evidence-files:start -->

Evidence files

Run evidence for this release is stored in the repository under evidence/:

• evidence/zaya_qlora_eval_result_zaya1-8b-coding-qlora-eval_release_summary.json

These files are compact local/Kaggle run artifacts used to document training, evaluation, merge, or quantization evidence for this model family.

<!-- evidence-files:end -->