ExternalVendorCode/LocalAI
1
0
Fork 0
mirror of https://github.com/mudler/LocalAI.git synced 2025-04-15 06:56:47 +00:00
Code Issues Actions 13 Packages Projects Releases Wiki Activity

chore(model gallery): add arliai_qwq-32b-arliai-rpr-v (#5137)

Browse Source 
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
This commit is contained in:
Ettore Di Giacinto 2025-04-08 09:40:26 +02:00 committed by GitHub
parent a6f0bb410f
commit 547d322b28
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 25 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
gallery/index.yaml
View File

@ -5684,6 +5684,31 @@
- filename: open-thoughts_OpenThinker2-7B-Q4_K_M.gguf
sha256: 481d785047d66ae2eeaf14650a9e659ec4f7766a6414b6c7e92854c944201734
uri: huggingface://bartowski/open-thoughts_OpenThinker2-7B-GGUF/open-thoughts_OpenThinker2-7B-Q4_K_M.gguf
- !!merge <<: *qwen25
name: "arliai_qwq-32b-arliai-rpr-v1"
icon: https://cdn-uploads.huggingface.co/production/uploads/6625f4a8a8d1362ebcc3851a/albSlnUy9dPVGVuLlsBua.jpeg
urls:
- https://huggingface.co/ArliAI/QwQ-32B-ArliAI-RpR-v1
- https://huggingface.co/bartowski/ArliAI_QwQ-32B-ArliAI-RpR-v1-GGUF
description: |
RpR (RolePlay with Reasoning) is a new series of models from ArliAI. This series builds directly upon the successful dataset curation methodology and training methods developed for the RPMax series.
RpR models use the same curated, deduplicated RP and creative writing dataset used for RPMax, with a focus on variety to ensure high creativity and minimize cross-context repetition. Users familiar with RPMax will recognize the unique, non-repetitive writing style unlike other finetuned-for-RP models.
With the release of QwQ as the first high performing open-source reasoning model that can be easily trained, it was clear that the available instruct and creative writing reasoning datasets contains only one response per example. This is type of single response dataset used for training reasoning models causes degraded output quality in long multi-turn chats. Which is why Arli AI decided to create a real RP model capable of long multi-turn chat with reasoning.
In order to create RpR, we first had to actually create the reasoning RP dataset by re-processing our existing known-good RPMax dataset into a reasoning dataset. This was possible by using the base QwQ Instruct model itself to create the reasoning process for every turn in the RPMax dataset conversation examples, which is then further refined in order to make sure the reasoning is in-line with the actual response examples from the dataset.
Another important thing to get right is to make sure the model is trained on examples that present reasoning blocks in the same way as it encounters it during inference. Which is, never seeing the reasoning blocks in it's context. In order to do this, the training run was completed using axolotl with manual template-free segments dataset in order to make sure that the model is never trained to see the reasoning block in the context. Just like how the model will be used during inference time.
The result of training QwQ on this dataset with this method are consistently coherent and interesting outputs even in long multi-turn RP chats. This is as far as we know the first true correctly-trained reasoning model trained for RP and creative writing.
overrides:
parameters:
model: ArliAI_QwQ-32B-ArliAI-RpR-v1-Q4_K_M.gguf
files:
- filename: ArliAI_QwQ-32B-ArliAI-RpR-v1-Q4_K_M.gguf
sha256: b0f2ca8f62a5d021e20db40608a109713e9d23e75b68b3b71b7654c04d596dcf
uri: huggingface://bartowski/ArliAI_QwQ-32B-ArliAI-RpR-v1-GGUF/ArliAI_QwQ-32B-ArliAI-RpR-v1-Q4_K_M.gguf
- &llama31
url: "github:mudler/LocalAI/gallery/llama3.1-instruct.yaml@master" ## LLama3.1
icon: https://avatars.githubusercontent.com/u/153379578