41 lines
1.3 KiB
Markdown
41 lines
1.3 KiB
Markdown
---
|
|
license: apache-2.0
|
|
base_model: veyra-ai/veyra2-30m-base-2b-tokens
|
|
tags:
|
|
- gguf
|
|
- llama.cpp
|
|
- text-generation
|
|
- veyra
|
|
- small-language-model
|
|
---
|
|
|
|
# Veyra2 30M Base 2B Tokens GGUF
|
|
|
|
GGUF conversions of [`veyra-ai/veyra2-30m-base-2b-tokens`](https://huggingface.co/veyra-ai/veyra2-30m-base-2b-tokens) for llama.cpp-compatible runtimes.
|
|
|
|
## Files
|
|
|
|
- `veyra2-30m-base-2b-tokens-f16.gguf`
|
|
- `veyra2-30m-base-2b-tokens-Q8_0.gguf`
|
|
- `veyra2-30m-base-2b-tokens-Q6_K.gguf`
|
|
- `veyra2-30m-base-2b-tokens-Q5_K_M.gguf`
|
|
- `veyra2-30m-base-2b-tokens-Q4_K_M.gguf`
|
|
- `veyra2-30m-base-2b-tokens-Q3_K_M.gguf`
|
|
- `veyra2-30m-base-2b-tokens-Q2_K.gguf`
|
|
|
|
## Notes
|
|
|
|
These GGUFs were converted with `tokenizer.ggml.add_eos_token=false`, because the original conversion metadata appended EOS to prompts, causing llama.cpp continuation prompts to behave like a new document after the prompt. If you are using something like LM Studio you can set your chat template to ChatML to talk to Veyra.
|
|
|
|
Recommended starting points:
|
|
|
|
- Best quality: `Q8_0` or `f16`
|
|
- Balanced: `Q5_K_M` or `Q6_K`
|
|
- Smallest experimental: `Q2_K` / `Q3_K_M`
|
|
|
|
Recommended settings:
|
|
|
|
- Temperature: 0.8
|
|
- Repeat Penalty: 1.2
|
|
|
|
**Note:** For lower quants I would recommend using higher temperature to avoid repetition over repeat penalty. |