Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: aimeri/spoomplesmaxx-mockingbird-36B
|
| 4 |
+
library_name: gguf
|
| 5 |
+
pipeline_tag: text-generation
|
| 6 |
+
tags:
|
| 7 |
+
- roleplay
|
| 8 |
+
- creative-writing
|
| 9 |
+
language:
|
| 10 |
+
- en
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# spoomplesmaxx-mockingbird-36B — GGUF (static)
|
| 14 |
+
|
| 15 |
+
Static GGUF quants of
|
| 16 |
+
[spoomplesmaxx-mockingbird-36B](https://huggingface.co/aimeri/spoomplesmaxx-mockingbird-36B),
|
| 17 |
+
the first of the mimids. Weighted/imatrix quants (calibrated on the model's
|
| 18 |
+
own training corpus) live in
|
| 19 |
+
[-i1-GGUF](https://huggingface.co/aimeri/spoomplesmaxx-mockingbird-36B-i1-GGUF);
|
| 20 |
+
prefer those at 3–4 bit if your runtime supports them.
|
| 21 |
+
|
| 22 |
+
| Quant | Size | Notes |
|
| 23 |
+
|---|---|---|
|
| 24 |
+
| Q3_K_M | ~18 GB | the 18GB target; fine with a good card |
|
| 25 |
+
| Q4_K_S | ~21 GB | |
|
| 26 |
+
| Q4_K_M | ~22 GB | recommended balance |
|
| 27 |
+
| Q5_K_M | ~26 GB | closest to bf16 behavior |
|
| 28 |
+
|
| 29 |
+
The seed-native chat template is embedded in the GGUF metadata — llama.cpp,
|
| 30 |
+
koboldcpp, and LM Studio pick it up automatically.
|
| 31 |
+
|
| 32 |
+
## Sampling — read this part
|
| 33 |
+
|
| 34 |
+
```
|
| 35 |
+
temperature 1.0 · top_p 0.9 · repeat_penalty 1.0 (OFF)
|
| 36 |
+
```
|
| 37 |
+
|
| 38 |
+
> **⚠ Never use repetition, presence, or frequency penalties.**
|
| 39 |
+
> The template ends every message with `<seed:eos>`; context-wide penalties
|
| 40 |
+
> suppress that token, the model stops ending its turns, and generation
|
| 41 |
+
> degenerates into the base model's untrained Chinese vocabulary. Many
|
| 42 |
+
> frontend presets default repeat_penalty to 1.05–1.1 — set it back to 1.0.
|
| 43 |
+
> Use DRY or XTC if you want extra anti-repetition; both leave special
|
| 44 |
+
> tokens alone.
|
| 45 |
+
|
| 46 |
+
Usable temperature window is ~0.95–1.05: lower loops verbatim, higher frays.
|
| 47 |
+
Full details, corpus notes, and training story on the
|
| 48 |
+
[main model card](https://huggingface.co/aimeri/spoomplesmaxx-mockingbird-36B).
|
| 49 |
+
|
| 50 |
+
*mimids 01 · Apache 2.0*
|