--- license: other license_name: yi-license license_link: LICENSE datasets: - adamo1139/rawrr_v2-2_stage1 - adamo1139/HESOYAM_v0.2 --- ## Known Issues There's something weird going on with tokenizer. EXL2 quant works fine in ooba but not in exui. BNB 4-bit quant works fine in ooba. For best results, use ooba with BOS token being inserted, repp 1.05 and probably exllamav2_HF loader over exllamav2 ## Model Description Have you ever wanted a sandbox for text-based social media? A place where you can bully a person, throw arguments or attack someone without any kind of actual harm being done and without any repercussions? All of it fully local, so nobody but you will ever know? No? Well, HESOYAM kinda can do that, but it's not exactly a bully similator, that's just one of ways you could use it. Specify a place on the internet that you want to be in the system prompt and then start a discussion. Will it be engaging or will you be sucked into someone's depression? For now, probably the latter. Still, I had some insightful concrete useful discussions with this model, it's not all gptslopped fluff. It does have a lot of depressive negative tones though, so it might not be for everyone. To get this model, first, I fine-tuned Yi-34B-200K (xlctx, as in second version of 34B 200K model, not new 1.5) on [adamo1139/rawrr_v2-2_stage1](https://huggingface.co/datasets/adamo1139/rawrr_v2-2_stage1) to make it so that base model will forget it's AI assistant programming and behave like a completion model trained on raw corpus of internet. This was done using [ORPO](https://arxiv.org/abs/2403.07691) and [GaLore](https://arxiv.org/abs/2403.03507) - all of it handled by [Unsloth](https://github.com/unslothai/unsloth). I would say it's a moderately successful finetune, I plan to enhance rawrr dataset with richer data to make better finetunes of this kind in the future. Resulting adapter file can be found [here](https://huggingface.co/adamo1139/Yi-34B-200K-XLCTX-RAW-ORPO-0805-GaLore-PEFT) and FP16 model file for RAWrr ORPO finetune can be found [here](https://huggingface.co/adamo1139/Yi-34B-200K-XLCTX-RAW-ORPO-0805-GaLore). Once I had good base model, I fine-tuned it on [HESOYAM 0.2](https://huggingface.co/datasets/adamo1139/HESOYAM_v0.2) dataset. It's a collection of single turn conversations from around 10 subreddits and multi-turn conversations from board /x/. There's also pippa in there. All samples there have system prompts that should tell the model about where discussion is taking place, this will be useful when you will be deciding on where you want to have your sandbox discussion take place. Here, I used classic SFT with GaLore and Unsloth, I wanted to get some results quick so it's trained for just 0.4 epochs. Adapter after that part of fine-tuning can be found [here](https://huggingface.co/adamo1139/Yi-34B-200K-XLCTX-HESOYAM-RAW-0905-GaLore-PEFT). [Conversation samples](https://huggingface.co/datasets/adamo1139/misc/blob/main/benchmarks/yi-34b-200k-xlctx-hesoyam-raw-0905/hesoyam_0905_samples.txt) - I put in a seed prompt and let the model generate the rest of the conversation. [Results on my base benchmarks](https://huggingface.co/datasets/adamo1139/misc/blob/main/benchmarks/yi-34b-200k-xlctx-hesoyam-raw-0905/benchmark_prompts.txt) - Responses suggests it still has some general assistant capabilities. I don't really want that, maybe I should up the learning rate for next run so that it stays in character more. ## Prompt template It's chatml, like always. ``` <|im_start|>system A chat on subreddit /r/pcmasterrace.<|im_end|> <|im_start|>user {prompt}<|im_end|> <|im_start|>assistant ``` ## Quants I haven't done them yet. I will maybe upload one EXL2 quant. ## Intended uses & limitations Use is limited by Yi license. \ Some datasets that were used prohibit commercial use (no_robots with CC-BY-NC-4.0), so I think you should use non-commercially only, unless you know law better and think it doesn't matter. ## Credits Thanks to unsloth and huggingface team for providing software packages used during fine-tuning. \ Thanks to authors of ORPO and GaLore for their innovative fine-tuning strategies. \ Thanks to random people who post datasets on hf, you rock! [made with Unsloth](https://github.com/unslothai/unsloth)