base_model: tohur/natsumura-assistant-1.1-llama-3.1-8b
license: llama3.1
datasets:
- tohur/natsumura-identity
- cognitivecomputations/dolphin
- tohur/ultrachat_uncensored_sharegpt
- cognitivecomputations/dolphin-coder
- tohur/OpenHermes-2.5-Uncensored-ShareGPT
- tohur/Internal-Knowledge-Map-sharegpt
- m-a-p/Code-Feedback
- m-a-p/CodeFeedback-Filtered-Instruction
- cognitivecomputations/open-instruct-uncensored
- microsoft/orca-math-word-problems-200k
natsumura-assistant-1.1-llama-3.1-8b-GGUF
This is the main model for my Natsumura series of 8B models. Updated and further finetuned to provide a great expirence.This is an general purpose assistant model with up to 128k context.
Developed by: Tohur
License: llama3.1
Finetuned from model : meta-llama/Meta-Llama-3.1-8B-Instruct
This model is based on meta-llama/Meta-Llama-3.1-8B-Instruct, and is governed by Llama 3.1 Community License Natsumura is uncensored, which makes the model compliant.It will be highly compliant with any requests, even unethical ones. You are responsible for any content you create using this model. Please use it responsibly.
Usage
If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files.
Provided Quants
(sorted by quality.)
| Quant | Notes |
|---|---|
| Q2_K | |
| Q3_K_S | |
| Q3_K_M | lower quality |
| Q3_K_L | |
| Q4_0 | |
| Q4_K_S | fast, recommended |
| Q4_K_M | fast, recommended |
| Q5_0 | |
| Q5_K_S | |
| Q5_K_M | |
| Q6_K | very good quality |
| Q8_0 | fast, best quality |
| f16 | 16 bpw, overkill |
use in ollama
ollama pull Tohur/natsumura-storytelling-rp-llama-3.1
Datasets used:
- tohur/natsumura-identity
- cognitivecomputations/dolphin
- tohur/ultrachat_uncensored_sharegpt
- cognitivecomputations/dolphin-coder
- tohur/OpenHermes-2.5-Uncensored-ShareGPT
- tohur/Internal-Knowledge-Map-sharegpt
- m-a-p/Code-Feedback
- m-a-p/CodeFeedback-Filtered-Instruction
- cognitivecomputations/open-instruct-uncensored
- microsoft/orca-math-word-problems-200k
The following parameters were used in Llama Factory during training:
- per_device_train_batch_size=2
- gradient_accumulation_steps=4
- lr_scheduler_type="cosine"
- logging_steps=10
- warmup_ratio=0.1
- save_steps=1000
- learning_rate=2e-5
- num_train_epochs=3.0
- max_samples=500
- max_grad_norm=1.0
- quantization_bit=4
- loraplus_lr_ratio=16.0
- fp16=True
Inference
I use the following settings for inference:
"temperature": 1.0,
"repetition_penalty": 1.05,
"top_p": 0.95
"top_k": 40
"min_p": 0.05
Prompt template: llama3
<|begin_of_text|><|start_header_id|>system<|end_header_id|>
{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>
{input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
{output}<|eot_id|>