README.md

metadata

license: other
library_name: transformers
tags:
  - finetune
  - synthetic data
  - custom_code
  - qwen2
  - COT
datasets:
  - kaist-ai/CoT-Collection
license_name: tongyi-qianwen-research
license_link: https://huggingface.co/Qwen/Qwen1.5-1.8B-Chat/raw/main/LICENSE
model-index:
  - name: Reyna-CoT-4B-v0.1
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: AI2 Reasoning Challenge (25-Shot)
          type: ai2_arc
          config: ARC-Challenge
          split: test
          args:
            num_few_shot: 25
        metrics:
          - type: acc_norm
            value: 44.71
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aloobun/Reyna-CoT-4B-v0.1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: HellaSwag (10-Shot)
          type: hellaswag
          split: validation
          args:
            num_few_shot: 10
        metrics:
          - type: acc_norm
            value: 71.12
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aloobun/Reyna-CoT-4B-v0.1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU (5-Shot)
          type: cais/mmlu
          config: all
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 55.9
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aloobun/Reyna-CoT-4B-v0.1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: TruthfulQA (0-shot)
          type: truthful_qa
          config: multiple_choice
          split: validation
          args:
            num_few_shot: 0
        metrics:
          - type: mc2
            value: 43.09
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aloobun/Reyna-CoT-4B-v0.1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: Winogrande (5-shot)
          type: winogrande
          config: winogrande_xl
          split: validation
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 67.72
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aloobun/Reyna-CoT-4B-v0.1
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GSM8k (5-shot)
          type: gsm8k
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 16.91
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=aloobun/Reyna-CoT-4B-v0.1
          name: Open LLM Leaderboard

• Finetuned Qwen/Qwen1.5-4B, on variety of CoT tasks including Reasoning, Closed Book Question Answering, Ethics, and more.
• Datasets : Curated from - kaist-ai/CoT-Collection, euclaise/TinyCoT and a very small subset from teknium/OpenHermes-2.5.
• This marks the fourth model in this series. This experiment aims to improve Chain of Thought (CoT) capabilities on smaller language models.
• I may rerun the finetuning experiment(with a more balanced dataset), using an iterative rationale-bootstrapping procedure inspired by euclaise/Memphis-CoT-3B.
• Hyperparameter: adamw with eps of 1e-8, cosine decay with 20% warmup, lr=2e-5

Benchamrks:

WIP

Example:

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer, StoppingCriteria
import torch

class MyStoppingCriteria(StoppingCriteria):
  def __init__(self, target_sequence, prompt):
    self.target_sequence = target_sequence
    self.prompt=prompt

  def __call__(self, input_ids, scores, **kwargs):
    generated_text = tokenizer.decode(input_ids[0])
    generated_text = generated_text.replace(self.prompt,'')
    if self.target_sequence in generated_text:
        return True 
    return False 

  def __len__(self):
    return 1

  def __iter__(self):
    yield self

modelpath="aloobun/Reyna-CoT-4B-v0.1"

model = AutoModelForCausalLM.from_pretrained(
    modelpath,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
    trust_remote_code=True,       
)

tokenizer = AutoTokenizer.from_pretrained(
    modelpath,
    trust_remote_code=True,      
    use_fast=False,
)

prompt = "Avery opens a flower shop. She ties 8 bunches of flowers with 9 flowers in each bunch. How many bunches would she have if she put 12 flowers in each bunch instead?\n"

encoded_input = tokenizer(prompt, return_tensors='pt')
input_ids=encoded_input['input_ids'].cuda()
streamer = TextStreamer(tokenizer=tokenizer, skip_prompt=True)
op = model.generate(
    input_ids,
    streamer=streamer,
    pad_token_id=tokenizer.eos_token_id,
    do_sample=True,
    temperature=0.6,
    top_p=0.8,
    max_new_tokens=512,
    stopping_criteria=MyStoppingCriteria("<|endoftext|>", prompt)
)

Output:

She would have 8 x 9 = 72 flowers in total. She would have 72 / 12 = 6 bunches of flowers with 12 flowers in each bunch. Therefore, the answer is 6.<|endoftext|>

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	49.91
AI2 Reasoning Challenge (25-Shot)	44.71
HellaSwag (10-Shot)	71.12
MMLU (5-Shot)	55.90
TruthfulQA (0-shot)	43.09
Winogrande (5-shot)	67.72
GSM8k (5-shot)	16.91