--- license: apache-2.0 datasets: - adamo1139/toxic-dpo-natural-v5 - adamo1139/AEZAKMI_v3-7 - adamo1139/rawrr_v2-2_stage1 --- ## Model description Yi-34B 200K XLCTX base model fine-tuned on adamo1139/rawrr_v2-2_stage1 (DPO), adamo1139/AEZAKMI_v3-7 (SFT) and adamo1139/toxic-dpo-natural-v5 (ORPO) datasets. Training took around 7 (DPO) + 13 (SFT) + 3 (ORPO) = 23 hours total on RTX 3090 Ti, all finetuning was done locally. This is excluding failed attempts and issues I had with merging script, that basically made me run DPO and SFT stages 2 times over because I thought that my LoRAs were broken, but it turned out to be some bug with new transformers/peft versions. This model is tuned to use more natural language and also be very uncensored. Say goodbye to "It's important to remember"! \ Prompt format is standard chatml. Don't expect it to be good at math, riddles or be crazy smart. My end goal with AEZAKMI is to create a cozy free chatbot. Cost of this fine-tune is about $5-$10 in electricity. Base model used for fine-tuning was Yi-34B-200K model shared by 01.ai, the newer version that has improved long context needle in a haystack retrieval. They didn't give it a new name, giving it numbers would mess up AEZAKMI naming scheme by adding a second number, so I will be calling it XLCTX. [You can see examples of responses to various prompts here (loaded with transformers load_in_4bit)](https://huggingface.co/datasets/adamo1139/misc/blob/main/benchmarks/yi-34b-200k-xlctx-aezakmi-raw-toxic-natural-orpo-0205/benchmark_prompts.txt) I had to lower max_positional_embeddings in config.json and model_max_length for training to start, otherwise I was OOMing straight away. This attempt had both max_position_embeddings and model_max_length set to 4096, which worked perfectly fine. I then reversed this to 200000 once I was uploading it. I think it should keep long context capabilities of the base model should be present here. If you want to see training scripts, let me know and I will upload them. LoRAs are uploaded [here adamo1139/yi-34b-200k-xlctx-aezakmi-raw-toxic-dpo-sft-orpo-lora-0205](https://huggingface.co/adamo1139/yi-34b-200k-xlctx-aezakmi-raw-toxic-dpo-sft-orpo-lora-0205) ## Quants! EXL2 quant coming soon, I plan to make and upload something around 4.65bpw, it should work nicely with q4 cache in exllama2 ## Prompt Format I recommend using ChatML format, as this was used during fine-tune. \ Here's a prompt format you should use, you can set a different system message, model was trained on SystemChat dataset, so it should respect system prompts fine. ``` <|im_start|>system A chat.<|im_end|> <|im_start|>user {prompt}<|im_end|> <|im_start|>assistant ``` ## Intended uses & limitations Use is limited by apache-2.0 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. ## Known Issues I haven't found any yet. ## Credits Thanks to unsloth and huggingface team for providing software packages used during fine-tuning. \ Thanks to Jon Durbin, abacusai, huggingface, sandex, NobodyExistsOnTheInternet, Nous-Research, lmsys, PygmalionAI for open sourcing datasets I included in the AEZAKMI dataset. \ AEZAKMI is basically a mix of open source datasets I found on HF, so without them this would not be possible at all. [made with Unsloth](https://github.com/unslothai/unsloth)