39 lines
1.0 KiB
Markdown
39 lines
1.0 KiB
Markdown
---
|
|
license: apache-2.0
|
|
base_model: ATH-MaaS/Marco-Nano-Instruct
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- reap
|
|
- pruning
|
|
- qwen3_moe
|
|
- moe
|
|
- safetensors
|
|
---
|
|
|
|
# Marco-Nano-Instruct REAP Pruned 0.25
|
|
|
|
This is a pruned derivative of `ATH-MaaS/Marco-Nano-Instruct`, produced with
|
|
`CerebrasResearch/reap` layerwise REAP pruning in a Codex-assisted experiment
|
|
(GPT5.5 xhigh based).
|
|
|
|
## Details
|
|
|
|
- Pruning method: REAP layerwise pruning
|
|
- Requested compression ratio: `0.25`
|
|
- Experts retained: `174`
|
|
- Experts per token: `8`
|
|
- Calibration dataset: `theblackcat102/evol-codealpaca-v1`
|
|
- `model_max_length`: `2048`
|
|
- `batches_per_category`: `128`
|
|
- `batch_size`: `1`
|
|
- `batch_group_size`: `8`
|
|
- Format: safetensors
|
|
|
|
## Notes
|
|
|
|
This checkpoint was created for a low-resource pruning experiment on a 6GB VRAM
|
|
environment. Benchmark evaluation was not run. A short prompt smoke check showed
|
|
usable English output, but Japanese and Chinese outputs may still show
|
|
repetition, self-continuation, or wording artifacts.
|