Commit
•
87553b5
1
Parent(s):
ccc2a35
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
---
|
| 3 |
+
|
| 4 |
+
library_name: transformers
|
| 5 |
+
license: apache-2.0
|
| 6 |
+
base_model: Heralax/philosophy-llm-mistral-pretrain
|
| 7 |
+
tags:
|
| 8 |
+
- generated_from_trainer
|
| 9 |
+
model-index:
|
| 10 |
+
- name: philosophy-hardcore-pretraining
|
| 11 |
+
results: []
|
| 12 |
+
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
[](https://hf.co/QuantFactory)
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
# QuantFactory/philosophy-mistral-GGUF
|
| 19 |
+
This is quantized version of [Heralax/philosophy-mistral](https://huggingface.co/Heralax/philosophy-mistral) created using llama.cpp
|
| 20 |
+
|
| 21 |
+
# Original Model Card
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
# Philosophy LLM
|
| 25 |
+
|
| 26 |
+
I would've trained this on Phi so I could've called it Phi-losophy if I had thought of that joke before kicking off the run. Oh well.
|
| 27 |
+
It's trained on Mistral instead. That's a Mist opportunity right there.
|
| 28 |
+
|
| 29 |
+
This is a narrow domain-expert LLM trained on the top 5 books on Gutenberg:
|
| 30 |
+
|
| 31 |
+
- The Problems of Philosophy (Bertrand Russell)
|
| 32 |
+
- Beyond Good and Evil (Nietzsche)
|
| 33 |
+
- Thus Spake Zarathustra: A Book for All and None (Nietzsche)
|
| 34 |
+
- The Prince (Machiavelli)
|
| 35 |
+
- Second Treatise of Government
|
| 36 |
+
|
| 37 |
+
It's meant to be an interesting novelty, showing off training on a specific domain. It has some quirks. Namely:
|
| 38 |
+
|
| 39 |
+
1. It seems to have memorized the training data very well. Ask a question that exists in the training data, with temp 0, and it will usually give you back the exact response word-for-word. This means that, on the subjects covered by its data, it will be very knowledgeable.
|
| 40 |
+
2. I forgot to include any generalist instruct data, so it's... not stupid, at least not particularly stupid by 7b standards, but it is very much limited to QA.
|
| 41 |
+
3. It's much less fluffy and wasteful with its responses than previous Augmentoolkit domain expert models, due to using a new dataset setting. This tends to make it respond with less detail, but it also may remember stuff better and get to the point easier.
|
| 42 |
+
|
| 43 |
+
Some example chats (blame LM studio for not hiding the stop token):
|
| 44 |
+
|
| 45 |
+
Asking stuff from the training data:
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
Asking a question directly from the training data and one I came up with on the spot.
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
Some things that are kinda funny but also show off the drawback of not using any generalist data:
|
| 52 |
+
)
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
## Training procedure
|
| 57 |
+
|
| 58 |
+
### Training hyperparameters
|
| 59 |
+
|
| 60 |
+
The following hyperparameters were used during training:
|
| 61 |
+
- learning_rate: 2e-05
|
| 62 |
+
- train_batch_size: 2
|
| 63 |
+
- eval_batch_size: 1
|
| 64 |
+
- seed: 42
|
| 65 |
+
- distributed_type: multi-GPU
|
| 66 |
+
- num_devices: 6
|
| 67 |
+
- gradient_accumulation_steps: 6
|
| 68 |
+
- total_train_batch_size: 72
|
| 69 |
+
- total_eval_batch_size: 6
|
| 70 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
| 71 |
+
- lr_scheduler_type: cosine
|
| 72 |
+
- lr_scheduler_warmup_steps: 136
|
| 73 |
+
- num_epochs: 6
|
| 74 |
+
|
| 75 |
+
### Framework versions
|
| 76 |
+
|
| 77 |
+
- Transformers 4.45.0.dev0
|
| 78 |
+
- Pytorch 2.3.1+cu121
|
| 79 |
+
- Datasets 2.21.0
|
| 80 |
+
- Tokenizers 0.19.1
|
| 81 |
+
|