Update the model weights

README.md

@@ -28,52 +28,20 @@ prompt_template: >
 
 This model is a version of [KoboldAI/LLaMA2-13B-Psyfighter2](https://huggingface.co/KoboldAI/LLaMA2-13B-Psyfighter2) finetuned to better understand vore context. The primary purpose of this model is to be a storywriting assistant, a conversational model in a chat, and an interactive choose-your-own-adventure text game.
 
-The Adventure Mode is still work in progress, and will be the focus of the future updates.
+The Adventure Mode is still work in progress, and will be added later.
 
-This is the FP16-precision version of the model for merging and fine-tuning. For using the model, download the quantized version here instead: [SnakyMcSnekFace/Psyfighter2-13B-vore-GGUF](https://huggingface.co/SnakyMcSnekFace/Psyfighter2-13B-vore-GGUF)
+This is the FP16-precision version of the model for merging and fine-tuning. **For using the model, please see the quantized version and the instructions here: [SnakyMcSnekFace/Psyfighter2-13B-vore-GGUF](https://huggingface.co/SnakyMcSnekFace/Psyfighter2-13B-vore-GGUF)**
 
 ## Model Details
 
-### Model Description
-
 The model behaves similarly to `KoboldAI/LLaMA2-13B-Psyfighter2`, which it was derived from. Please [see the README.md here](https://huggingface.co/KoboldAI/LLaMA2-13B-Psyfighter2/blob/main/README.md) to learn more.
 
-This model was fine-tuned on ~55 MiB of free-form text, containing stories focused around the vore theme. As a result, it has a strong vorny bias.
-
-## How to Get Started with the Model
-
-The model can be used with any AI chatbots and front-ends designed to work with `.gguf` models. The model fits fully into 8GB VRAM, but can also run with degraded performance on smaller graphics cards.
-
-Similarly to the base model, the less prompt the model receives, the more creative is the output. For example, the writing assistant will generate an entire story when prompted with only 2-3 words.
-
-In the chat mode, if the conversation is not going where you would like it to go, edit the model's output and let it continue generation. The model will also match the style of the conversation.
-
-### Koboldcpp Colab Notebook
-
-The easiest way to try out the model is [Koboldcpp Colab Notebook](https://colab.research.google.com/github/lostruins/koboldcpp/blob/concedo/colab.ipynb). This method doesn't require you to have a powerful graphics card.
-
-- Open the notebook
-- Paste the model URL into the field: `https://huggingface.co/SnakyMcSnekFace/Psyfighter2-13B-vore-GGUF/resolve/main/Psyfighter2-13B-vore.Q4_K_M.gguf`
-- Start the notebook, wait for the URL to CloudFlare tunnel to appear at the bottom and click it
-- Use the model as a writing assistant
-- Try an adventure from [https://aetherroom.club/](https://aetherroom.club/). Adventure mode is still work-in-progress, but it's getting there.
-
-### Backyard AI 
-
-Another convenient way to use the model is [Backyard AI](https://backyard.ai/) application, which allows you to run the model locally on your computer. You'll need a graphics card with at least 8GB VRAM to use the model comfortably.
-
-#### Download directly from HuggingFace (beta)
-
-In the left panel, click `Manage Models`, then select `Hugging face models`. Paste `https://huggingface.co/SnakyMcSnekFace/Psyfighter2-13B-vore-GGUF` into the text field and press `Fetch Models`. Click `Download` button to the next to the model format. Once the model is downloaded, you can select it in your character card or set it as a default model.
-
-#### Download manually
+### Updates
 
-Download the [Psyfighter2-13B-vore.Q4_K_M.gguf](https://huggingface.co/SnakyMcSnekFace/Psyfighter2-13B-vore-GGUF/resolve/main/Psyfighter2-13B-vore.Q4_K_M.gguf) file into `%appdata%\faraday\models` folder on your computer. The model should appear in `Manage Models` menu under `Downloaded Models`. You can then select it in your character card or set it as a default model.
-
-### Model updates
-
-- 04/13/2024 - uploaded the first version of the model
+- 06/02/2024 - fixed errors in training and merging, significantly improving the overall prose quality
 - 05/25/2024 - updated training process, making the model more coherent and improving the writing quality
+- 04/13/2024 - uploaded the first version of the model
+
 
 ## Bias, Risks, and Limitations
 
@@ -81,14 +49,15 @@ By design, this model has a strong vorny bias. It's not intended for use by anyo
 
 ## Training Details
 
-This model was fine-tuned on free-form text comprised of stories focused around the vore theme using [rank-stabilized](https://arxiv.org/abs/2312.03732) [QLoRA adapter](https://arxiv.org/abs/2305.14314) [QLoRA method](https://arxiv.org/abs/2305.14314). The resulting adapter was merged into the FP16 precision base model. The quantized version of the model was prepared using [llama.cpp](https://github.com/ggerganov/llama.cpp).
+The model was fine-tuned using a [rank-stabilized](https://arxiv.org/abs/2312.03732) [QLoRA adapter](https://arxiv.org/abs/2305.14314). Training was performed using [Unsloth AI](https://github.com/unslothai/unsloth) library on `Ubuntu 22.04.4 LTS` with `CUDA 12.1` and `Pytorch 2.3.0`.
 
-### Training Procedure
+The total training time on NVIDIA GeForce RTX 4060 Ti is about 24 hours.
 
-The model was fine-tuned with a [rank-stabilized](https://arxiv.org/abs/2312.03732) [QLoRA adapter](https://arxiv.org/abs/2305.14314) on NVIDIA GeForce RTX 4060 Ti over the span of ~24 hours. Training was performed using [Unsloth AI](https://github.com/unslothai/unsloth) library on `Ubuntu 22.04.4 LTS` with `CUDA 12.1` and `Pytorch 2.3.0`.
+After training, the adapter weights were merged into the dequantized model as described in [ChrisHayduk's GitHub gist](https://gist.github.com/ChrisHayduk/1a53463331f52dca205e55982baf9930).
 
+The quantized version of the model was prepared using [llama.cpp](https://github.com/ggerganov/llama.cpp).
 
-#### LoRa adapter configuration
+### LoRa adapter configuration
 
 - Rank: 128
 - Alpha: 16
@@ -96,10 +65,24 @@ The model was fine-tuned with a [rank-stabilized](https://arxiv.org/abs/2312.037
 - Target weights: `["q_proj", "k_proj", "o_proj", "gate_proj", "up_proj"]`,
 - `use_rslora=True`
 
+
+### Domain adaptation 
+
+The initial training phase consists of fine-tuning the adapter on ~55 MiB of free-form text that containing stories focused around the vore theme. The text is broken into paragraphs, which are aggregated into training samples of 4096 tokens or less, without crossing the document boundary. Each sample starts with BOS token (with its `attention_mask` set to 0), and ends in EOS token. The paragraph breaks are normalized to always consist of two line breaks.
+
+#### Dataset pre-processing
+
+The raw-text stories in dataset were edited as follows:
+
+- titles, foreword, tags, and anything not comprising the text of the story are removed
+- non-ascii characters and chapter separators are removed
+- stories mentioning underage personas in any context are deleted
+- names of private characters are randomized
+
 #### Training parameters
 
 - Max. sequence length: 4096 tokens
-- Samples per epoch: 3783
+- Samples per epoch: 5085
 - Number of epochs: 2
 - Learning rate: 1e-4
 - Warmup: 64 steps
@@ -107,21 +90,11 @@ The model was fine-tuned with a [rank-stabilized](https://arxiv.org/abs/2312.037
 - Batch size: 1
 - Gradient accumulation steps: 1
 
-#### Preprocessing
-
-The stories in dataset were pre-processed as follows:
 
-- titles, foreword, tags, and anything not comprising the text of the story was removed
-- non-ascii characters and character sequences serving as chapter separators were removed
-- any story mentioning underage personas in any context was removed from the dataset
-- names of private characters were replaced with randomized names across the dataset
+### Adventure mode SFT
 
-## Environmental Impact
+TBD
 
-Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+### Adventure mode KTO
 
-- **Hardware Type:** NVIDIA GeForce RTX 4060 Ti
-- **Hours used:** 24
-- **Cloud Provider:** N/A
-- **Compute Region:** US-East
-- **Carbon Emitted:** 0.83 kg CO2 eq.
+TBD

config.json

@@ -1,5 +1,5 @@
 {
-  "_name_or_path": "../../../models/KoboldAI_LLaMA2-13B-Psyfighter2/",
+  "_name_or_path": "/tmp/dequantized_Psyfighter2-13B/",
   "architectures": [
     "LlamaForCausalLM"
   ],

model-00001-of-00006.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e405b05644ace92bb9be26ecd41824a2a773a70fa19ee534c024c9e53dc3e8e6
+oid sha256:433625fd1da3ae8163cbb623be665f1b7f40252f8b7a6bbc0d77f74a88459c60
 size 4978265728

model-00002-of-00006.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5b9bec5b5e9b28a6f68ed964381ab07017a6dcd06c1507bf802d98b23af55a23
+oid sha256:2468fd858ecf4ec3c290c93f02b20c32729022772cacad75350e339266d829aa
 size 4970422160

model-00003-of-00006.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3a77527d52ba11b1dbd01c627ca81c28e2dbab2914cf6d6bf4de265a855f1a3e
+oid sha256:ac2e69f709b2f8f55774095f524f6d2cca4d0842aaf8d76e202cfc263c428f38
 size 4970422184

model-00004-of-00006.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:971e8f443e04c074ae416fbccceb49749608247e3893c2805bdc072b16bed550
+oid sha256:f12b3071b6006608aa1dac69ff465dbbb9d2ed1fb9989d02d77aece0e4402d14
 size 4933701432

model-00005-of-00006.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d05c74b09dc9e6f54fd13bec21f19704dfbb126484cab9115a3ca95ce3cd38e3
+oid sha256:50780887ba3c2602c8fc3ffb58c43c6144d42442bef75e71d33842daa0b117a9
 size 4933722144

model-00006-of-00006.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:85653cec6b1526afd2de089f710528bbd3e141ed91a9f3526816b05700a4d121
+oid sha256:ca60917d349df6204c8e6e4f890beacd837d1ad786190b8cef9f5d3266556c27
 size 1245236904