HF :
FP16
|
GGUF :
imatrix ·
static ·
Q6_K ·
Q4_K_M
|
EXL2 :
4 BPW ·
6 BPW ·
8 BPW
Thanks to SicariusSicariiStuff for the help with training &
mradermacher for the imatrix & static GGUFs, and
Statuo for the EXL2s.
This is a fix for the quantization issue in Firefly v0.1.
Firefly is a Mistral Small 22B finetune designed for creative writing and roleplay. The model is largely uncensored and should support context up to 32,768 tokens.
The model has been tested in various roleplay scenarios up to 16k context, as well as in a role as an assistant. It shows a broad competency & coherence across various scenarios.
Special thanks to SicariusSicariiStuff for bouncing ideas back & forth on training, and SytanSD for quants.
I appreciate all feedback on any of my models, you can use:
Your feedback is how I improve these models for future versions.
This model is extensively uncensored. It can generate explicit, disturbing or offensive responses. Use responsibly. I am not responsible for your use of this model.
This model is a finetune of Mistral Small 22B (2409) and usage must follow the terms of Mistral's license. By downloading this model, you agree not to use it for commercial purposes unless you have a valid Mistral commercial license. See the base model card for more details.
I'd recommend Mistral v2 & v3 prompting format:
<s>[INST] User message here.[/INST] Bot response here</s>[INST] User message 2 here.[/INST]
If you're using SillyTavern, make sure the story string is set correctly to Mistral v2 & v3 (not v3 Tekken):
[INST] {{#if system}}{{system}}
{{/if}}{{#if wiBefore}}{{wiBefore}}
{{/if}}{{#if description}}{{description}}
{{/if}}{{#if personality}}{{personality}}
{{/if}}{{#if scenario}}{{scenario}}
{{/if}}{{#if wiAfter}}{{wiAfter}}
{{/if}}{{#if persona}}{{persona}}
{{/if}}{{trim}}[/INST] Understood.</s>
The model seems very sensitive to wrong prompting formats.
I'm running the following sampler settings but this is an RC and they may not be optimal.
If you get completely incoherent responses, feel free to use these as a starting point.
High temperature settings (above 1) tend to create less coherent responses.
I started with a finetune of Mistral Small 22B which had been trained on the Gutenberg dataset: nbeerbower/Mistral-Small-Gutenberg-Doppel-22B.
The first stage of my training was a single epoch at low LR over a 474 million token text completion dataset.
I followed this up with a coherence, decensorship & roleplay finetune over a 172 million token instruct dataset over two epochs.
I did a slerp merge of epoch 1 into epoch 2 at a light weight which resolved the name-spelling issues on quantized versions of Firefly v0.1.
Total training time was about 32hrs on 4x Nvidia A100 80GB.
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_invisietch__MiS-Firefly-v0.2-22B)
| Metric | Value |
|---|---|
| Avg. | 26.65 |
| IFEval (0-Shot) | 53.71 |
| BBH (3-Shot) | 36.08 |
| MATH Lvl 5 (4-Shot) | 15.94 |
| GPQA (0-shot) | 7.27 |
| MuSR (0-shot) | 17.81 |
| MMLU-PRO (5-shot) | 29.11 |