jordiclive's picture
Update README.md
dd588a4
|
raw
history blame
5.62 kB
metadata
license: mit
datasets:
  - sahil2801/CodeAlpaca-20k
  - yahma/alpaca-cleaned
  - databricks/databricks-dolly-15k
  - OpenAssistant/oasst1
  - jeffwan/sharegpt_vicuna
  - qwedsacf/grade-school-math-instructions
  - vicgalle/alpaca-gpt4
language:
  - en
tags:
  - sft
pipeline_tag: text-generation
widget:
  - text: >-
      <|prompter|>What is a meme, and what's the history behind this
      word?</s><|assistant|>
  - text: <|prompter|>What's the Earth total population</s><|assistant|>
  - text: <|prompter|>Write a story about future of AI development</s><|assistant|>

LoRA Adapter for LLaMA 7B trained on more datasets than tloen/alpaca-lora-7b

This repo contains a low-rank adapter for LLaMA-7b fit on datasets part of the OpenAssistant project.

You can see sampling results here. Note the sampling params are not necessarily the optimum—they are OpenAssistant defaults for comparing models.

This version of the weights was trained with the following hyperparameters:

  • Epochs: 8
  • Batch size: 128
  • Max Length: 2048
  • Learning rate: 8e-6
  • Lora r: 16
  • Lora Alpha: 32
  • Lora target modules: q_proj, k_proj, v_proj, o_proj

The model was trained with flash attention and gradient checkpointing.

Dataset Details

  • dolly15k: val_split: 0.05 max_val_set: 300
  • oasst_export: lang: "bg,ca,cs,da,de,en,es,fr,hr,hu,it,nl,pl,pt,ro,ru,sl,sr,sv,uk" input_file_path: 2023-04-12_oasst_release_ready_synth.jsonl.gz val_split: 0.05
  • vicuna: val_split: 0.05 max_val_set: 800 fraction: 0.8
  • dolly15k: val_split: 0.05 max_val_set: 300
  • grade_school_math_instructions: val_split: 0.05
  • code_alpaca: val_split: 0.05 max_val_set: 250
  • alpaca_gpt4: val_split: 0.02 max_val_set: 250

Model Details

  • Developed as part of the OpenAssistant Project
  • Model type: PEFT Adapter for frozen LLaMA
  • Language: English

Prompting

Two special tokens are used to mark the beginning of user and assistant turns: <|prompter|> and <|assistant|>. Each turn ends with a <|endoftext|> token.

Input prompt example:

<|prompter|>What is a meme, and what's the history behind this word?</s><|assistant|>

The input ends with the <|assistant|> token to signal that the model should start generating the assistant reply.

Example Inference Code (Note several embeddings need to be loaded along with the LoRA weights), assumes on GPU and torch.float16:

from typing import List, NamedTuple

import torch
import transformers
from huggingface_hub import hf_hub_download
from peft import PeftModel
from transformers import GenerationConfig

device = "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = transformers.AutoTokenizer.from_pretrained("jordiclive/alpaca_gpt4-dolly_15k-vicuna-lora-7b")


model = transformers.AutoModelForCausalLM.from_pretrained(
    "decapoda-research/llama-7b-hf", torch_dtype=torch.float16
)  # Load Base Model
model.resize_token_embeddings(
    len(tokenizer)
)  # This model repo also contains several embeddings for special tokens that need to be loaded.

model.config.eos_token_id = tokenizer.eos_token_id
model.config.bos_token_id = tokenizer.bos_token_id
model.config.pad_token_id = tokenizer.pad_token_id

lora_weights = "jordiclive/alpaca_gpt4-dolly_15k-vicuna-lora-7b"
model = PeftModel.from_pretrained(
    model,
    lora_weights,
    torch_dtype=torch.float16,
)  # Load Lora model

model.eos_token_id = tokenizer.eos_token_id
filename = hf_hub_download("jordiclive/alpaca_gpt4-dolly_15k-vicuna-lora-7b", "extra_embeddings.pt")
embed_weights = torch.load(
    filename, map_location=torch.device("cuda" if torch.cuda.is_available() else "cpu")
)  # Load embeddings for special tokens
model.base_model.model.model.embed_tokens.weight[32000:, :] = embed_weights.to(
    model.base_model.model.model.embed_tokens.weight.dtype
).to(
    device
)  # Add special token embeddings


model = model.half().to(device)
generation_config = GenerationConfig(
    temperature=0.1,
    top_p=0.75,
    top_k=40,
    num_beams=4,
)


def format_system_prompt(prompt, eos_token="</s>"):
    return "{}{}{}{}".format(
        "<|prompter|>",
        prompt,
        eos_token,
        "<|assistant|>"
    )

def generate(prompt, generation_config=generation_config, max_new_tokens=2048, device=device):
    prompt = format_system_prompt(prompt)  # OpenAssistant Prompt Format expected
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
    with torch.no_grad():
        generation_output = model.generate(
            input_ids=input_ids,
            generation_config=generation_config,
            return_dict_in_generate=True,
            output_scores=True,
            max_new_tokens=max_new_tokens,
            eos_token_id=2,
        )
    s = generation_output.sequences[0]
    output = tokenizer.decode(s)
    print("Text generated:")
    print(output)
    return output


generate("What is a meme, and what's the history behind this word?")
generate("What's the Earth total population")
generate("Write a story about future of AI development")