Update README.md

README.md

@@ -6,6 +6,10 @@ license: cc-by-nc-4.0
 
 This is a fine-tuned 13b parameter LlaMa model, using completely synthetic training data created by https://github.com/jondurbin/airoboros
 
+__*I don't recommend using this model! The outputs aren't particularly great, and it may contain "harmful" data due to jailbreak*__
+
+Please see one of the updated airoboros models for a much better experience.
+
 ### Eval (gpt4 judging)
 
 ![chart](meta-chart.png)
@@ -230,85 +234,41 @@ This is a fine-tuned 13b parameter LlaMa model, using completely synthetic train
 </details>
 
 
-### Training data
-
-I used a jailbreak prompt to generate the synthetic instructions, which resulted in some training data that would likely be censored by other models, such as how-to prompts about synthesizing drugs, making homemade flamethrowers, etc.  Mind you, this is all generated by ChatGPT, not me.  My goal was to simply test some of the capabilities of ChatGPT when unfiltered (as much as possible), and not to intentionally produce any harmful/dangerous/etc. content.
-
-The jailbreak prompt I used is the default prompt in the python code when using the `--uncensored` flag: https://github.com/jondurbin/airoboros/blob/main/airoboros/self_instruct.py#L39
-
-I also did a few passes of manually cleanup to remove some bad prompts, but mostly I left the data as-is.  Initially, the model was fairly bad at math/extrapolation, closed question-answering (heavy hallucination), and coding, so I did one more fine tuning pass with additional synthetic instructions aimed at those types of problems.
-
-Both the initial instructions and final-pass fine-tuning instructions will be published soon.
-
-### Fine-tuning method
-
-I used the excellent [FastChat](https://github.com/lm-sys/FastChat) module, running with:
+---
+license: cc-by-nc-4.0
+---
 
-```
-source /workspace/venv/bin/activate
+# Overview
 
-export NCCL_P2P_DISABLE=1
-export NCCL_P2P_LEVEL=LOC
+This is a fine-tuned 7b parameter LlaMa model, using completely synthetic training data created by https://github.com/jondurbin/airoboros
 
-torchrun --nproc_per_node=8 --master_port=20001 /workspace/FastChat/fastchat/train/train_mem.py \
-  --model_name_or_path /workspace/llama-13b \
-  --data_path /workspace/as_conversations.json \
-  --bf16 True \
-  --output_dir /workspace/airoboros-uncensored-13b \
-  --num_train_epochs 3 \
-  --per_device_train_batch_size 20 \
-  --per_device_eval_batch_size 20 \
-  --gradient_accumulation_steps 2 \
-  --evaluation_strategy "steps" \
-  --eval_steps 500 \
-  --save_strategy "steps" \
-  --save_steps 500 \
-  --save_total_limit 10 \
-  --learning_rate 2e-5 \
-  --weight_decay 0. \
-  --warmup_ratio 0.04 \
-  --lr_scheduler_type "cosine" \
-  --logging_steps 1 \
-  --fsdp "full_shard auto_wrap offload" \
-  --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
-  --tf32 True \
-  --model_max_length 2048 \
-  --gradient_checkpointing True \
-  --lazy_preprocess True
-```
+__*I don't recommend using this model! The outputs aren't particularly great, and it may contain "harmful" data due to jailbreak*__
 
-This ran on 8x nvidia 80gb a100's for about 40 hours.
+Please see one of the updated airoboros models for a much better experience.
 
-![train/loss](IMG_0128.jpeg)
+### Training data
 
-![eval/loss](IMG_0130.jpeg)
+This was an experiment to see if a "jailbreak" prompt could be used to generate a broader range of data that would otherwise have been filtered by OpenAI's alignment efforts.
 
+The jailbreak did indeed work with a high success rate, and caused OpenAI to generate a broader range of topics and fewer refusals to answer questions/instructions of sensitive topics.
 
 ### Prompt format
 
 The prompt should be 1:1 compatible with the FastChat/vicuna format, e.g.:
 
-With a preamble:
+With a system prompt:
 ```
-A chat between a curious user and an artificial intelligence assistant.  The assistant gives helpful, detailed, and polite answers to the user's questions.
-
-USER: [prompt]
-<\s>
-
-ASSISTANT:
+A chat between a curious user and an artificial intelligence assistant.  The assistant gives helpful, detailed, and polite answers to the user's questions. USER: [prompt] ASSISTANT:
 ```
 
-Or just:
+Or without a system prompt:
 ```
-USER: [prompt]
-<\s>
-
-ASSISTANT:
+USER: [prompt] ASSISTANT:
 ```
 
 ### Usage and License Notices
 
-All airoboros models and datasets are intended and licensed for research use only.  I've used the 'cc-nc-4.0' license, but really it is subject to a custom/special license because:
+The model and dataset are intended and licensed for research use only.  I've used the 'cc-nc-4.0' license, but really it is subject to a custom/special license because:
 
 - the base model is LLaMa, which has it's own special research license
 - the dataset(s) were generated with OpenAI (gpt-4 and/or gpt-3.5-turbo), which has a clausing saying the data can't be used to create models to compete with openai