Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
base_model: Pageshift-Entertainment/pagestorm-research-preview-14b-full-book
|
| 6 |
+
base_model_relation: quantized
|
| 7 |
+
tags:
|
| 8 |
+
- gguf
|
| 9 |
+
- llama.cpp
|
| 10 |
+
- story-generation
|
| 11 |
+
- staged-generation
|
| 12 |
+
- full-book
|
| 13 |
+
- ministral3
|
| 14 |
+
pipeline_tag: text-generation
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# PageStorm Research Preview 14B Full Book — Q8_0 GGUF
|
| 18 |
+
|
| 19 |
+
Q8_0 GGUF quantization of
|
| 20 |
+
[Pageshift-Entertainment/pagestorm-research-preview-14b-full-book](https://huggingface.co/Pageshift-Entertainment/pagestorm-research-preview-14b-full-book),
|
| 21 |
+
a `ministral3` model trained to produce a full novel from a single prompt via a
|
| 22 |
+
staged generation pipeline.
|
| 23 |
+
|
| 24 |
+
## Files
|
| 25 |
+
- `pagestorm-research-preview-14b-full-book-Q8_0.gguf` (~14 GB)
|
| 26 |
+
|
| 27 |
+
## Requirements
|
| 28 |
+
- A llama.cpp build whose runtime supports the **`mistral3`** architecture
|
| 29 |
+
(`llm_build_mistral3` / `LLM_ARCH_MISTRAL3`). Older builds will fail to load it.
|
| 30 |
+
|
| 31 |
+
## Notes
|
| 32 |
+
- Converted with `convert_hf_to_gguf.py --outtype q8_0`. The source `config.json`
|
| 33 |
+
needed `original_max_position_embeddings` changed from `16384.0` to integer
|
| 34 |
+
`16384` so the converter could write the int rope KV field.
|
| 35 |
+
- The model uses a **staged** protocol with custom role headers
|
| 36 |
+
(`<|start_header_id|>…<|stop_header_id|>`) and `<|eot_id|>` as the stage stop
|
| 37 |
+
token — it is not a plain chat model. See the base model card and its
|
| 38 |
+
`story_stage_generation.py` for the prompt protocol.
|
| 39 |
+
- Native context is 262144; KV at that length is large — quantize the KV cache
|
| 40 |
+
(`--cache-type-k q8_0 --cache-type-v q8_0`) and/or cap `--ctx-size` to fit VRAM.
|
| 41 |
+
|
| 42 |
+
## Attribution
|
| 43 |
+
Base model © Pageshift Entertainment, Apache-2.0. This repo only redistributes a
|
| 44 |
+
quantized copy of those weights.
|