Finetunes Llama-13b+Alaca to solve problems via stepwise reasoning (OpenAI PRM800k dataset, or rather our postprocessed version, Birchlabs/openai-prm800k-solutions-only).
Model description
This is a fork of llama-13b + chansung/alpaca-lora-13b.
That is: we loaded Llama-13b, we applied Alpaca LoRA, expanded vocabulary, then QLoRA 4-bit finetuned from there.
Parts:
- base model
llama-13b - LoRA 0
chansung/alpaca-lora-13b - LoRA 1
Birchlabs/llama-13b-stepwise-adapter(this) - tokenizer
Birchlabs/llama-13b-stepwise-tokenizer - finetuned input/output embedding layers:
Birchlabs/llama-13b-stepwise-embeddings
Training
Trained using qlora.py from our stepwise branch of qlora.
Known-good as of commit 522d86b.
python -m qlora --model_name_or_path huggyllama/llama-13b --lora_name_or_path chansung/alpaca-lora-13b --dataset prm800k-solutions --dataset_format prm800k-solutions --bf16 --max_memory_MB 24000 --use_bos_token_in_prompt --truncate_toward_center --source_max_len 184 --target_max_len 998 --gradient_accumulation_steps 4 --per_device_train_batch_size 4 --per_device_eval_batch_size 4 --learning_rate 0.0002 --run_name 13b_alpaca_special_tokens_long --report_to wandb --save_steps 64 --save_total_limit 3 --max_steps 1664 --evaluation_strategy steps --eval_steps 64 --generate_steps 16 --register_process_supervision_tokens
Usage
You can load using evaluate.py from our stepwise branch of qlora.
Known-good as of commit 522d86b.
Download embed_tokens.pt and lm_head.pt from Birchlabs/llama-13b-stepwise-embeddings, then run evaluator like so:
python -m evaluate \
--model_name_or_path huggyllama/llama-13b \
--base_lora_model_name_or_path chansung/alpaca-lora-13b \
--tokenizer_model_name_or_path Birchlabs/llama-13b-stepwise-tokenizer \
--lora_model_name_or_path Birchlabs/llama-13b-stepwise-adapter \
--input_embedding_path embed_tokens.pt \
--output_embedding_path lm_head.pt \
--bf16 \
--use_bos_token_in_prompt \
--overrun_countermeasures False
Prompt style
The model is finetuned for instruction-response.
prompt = '<s>Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n### Instruction:\nHow do I feed a family of 5 without ingredients?\n\n### Response:\n<|step_start|>'
Example output:
$ How do I drink a matcha latte?To drink a matcha latte, I should first remove the matcha from the bag. <|step_end|> <|step_start|> I can do that by opening the bag and shaking out the matcha into a glass or bowl. <|step_end|> <|step_start|> Then, I should add some hot water into the glass or bowl and start stirring the matcha. <|step_end|> <|step_start|> This is to make the matcha easier to mix with the water. <|step_end|> <|step_start|> Next, I should add milk into the glass or bowl and continue stirring. <|step_end|> <|step_start|> This is to make the matcha and the water form a thick mixture. <|step_end|> <|step_start|> Finally, I should add ice cubes into the glass or bowl and stir again. <|step_end|> <|step_start|> This is to make the mixture even thicker and chilled. <|step_end|> <|step_start|> So, a matcha latte is a drink that is a mixture of matcha, water, and milk, with some ice cubes. <|step_end|> <|step_start|> <|answer_start|> To make a matcha latte, I should remove the matcha from the bag, add some hot water into a glass or bowl and stir, add milk into the glass or bowl and stir as well, and finally add ice cubes and stir again. <|step_end|> <|step_start|> <|answer_start|> To drink a matcha latte, I should remove
$ What is the third Fibonacci number?I know that the first two Fibonacci numbers are 1 and 1, since they are just the ones that appear in the Fibonacci sequence: 1, 1, 2, 3, 5, 8, 13, and so on, where each term is the sum of the previous two. <|step_end|> <|step_start|> To find the third number, I could try to figure out a pattern or a shortcut, but I'll just use the definition and do some calculations. <|step_end|> <|step_start|> The third Fibonacci number is the sum of the first two, so I have 1 + 1 = 2. <|step_end|> <|step_start|> <|answer_start|> 2 <|answer_end|>
$ How do I make a cup of tea?Ok, the first step is to open the cup of tea. <|step_end|> <|step_start|> Then the second step is to jump in the tea. <|step_end|> <|step_start|> And finally the third step is to drink the jump. <|step_end|> <|step_start|> Let's reverse the order of the steps so that they are more intuitive. <|step_end|> <|step_start|> The first step is to jump in the tea. <|step_end|> <|step_start|> Then the second step is to open the cup of tea. <|step_end|> <|step_start|> And finally the third step is to drink the jump. <|step_end|> <|step_start|> So, I think this is a good way to make a cup of tea. <|step_end|> <|step_start|> Yes, it is. <|answer_start|> How do I make a cup of tea? <|answer_end|>
$ How do I open a jar?To open a jar, I need to use a key that fits into the lock on the jar. <|step_end|> <|step_start|> I recognize that a common type of key is a skeleton key, which can open many different types of locks, because it has no ridges that would restrict its usage. <|step_end|> <|step_start|> To get a skeleton key, I can cut any old key into pieces and join them back together. <|step_end|> <|step_start|> Alternatively, I can use a special tool that can make cuts exactly perpendicular to each other, forming a key that opens any lock. <|step_end|> <|step_start|> Alternatively, I can use a special trick that can make any old key open the jar. <|step_end|> <|step_start|> This trick involves placing the key on top of the jar and banging it on a hard surface, such as the floor or a table. <|step_end|> <|step_start|> I need to make sure that the key and the jar are both upside-down, and that I am banging them as hard as I can. <|step_end|> <|step_start|> This will create a small hole in the glass on the jar, and then I can use the key to open it. <|step_end|> <|step_start|> I need to be careful not to get cut by the glass, and I need to check if the jar is empty first, otherwise I might poke my finger and
- Downloads last month
- 0