rocket-3B / README.md
pansophic's picture
Update README.md
2716ba4
|
raw
history blame
7.02 kB
metadata
model-index:
  - name: rocket-3b
    results: []
license: cc-by-sa-4.0
language:
  - en
base_model: stabilityai/stablelm-3b-4e1t
Rocket Logo

Rocket-3B 🦝

Rocket 🦝 is a 3 billion large language model that was trained on a mix of publicly available datasets using Direct Preference Optimization (DPO). The prompt format used is ChatML.

Model description

  • Model type: A 3B parameter GPT-like model fine-tuned on a mix of publicly available datasets using DPO.
  • Language(s) (NLP): Primarily English
  • License: CC-BY-SA-4.0
  • Finetuned from model: Stability AI

Performance

Despite its compact dimensions, the model achieves outstanding scores in both MT-Bench MT-Bench and AlpacaEval benchmarks, surpassing the performance of considerably larger models.

Model Size Alignment MT-Bench (score) AlpacaEval (win rate %)
StableLM-Tuned-α 🦜 7B SFT 2.75 -
MPT-Chat 7B SFT 5.42 -
Falcon-Instruct πŸ¦… 40B SFT 5.17 45.71
Orca-2 13B SFT 6.15 -
Xwin-LMv0.1 7B PPO 6.19 87.83
Llama2-Chat πŸ¦™ 7B RLHF 6.26 71.37
TÜLU 2 🐫 7B DPO 6.27 85.1
Guanaco πŸ¦™ 65B SFT 6.41 71.80
Rocket 🦝 3B DPO 6.56 79.75
Llama2-Chat πŸ¦™ 13B RLHF 6.65 81.09
Zephyr-7b-Ξ± πŸͺ 7B DPO 6.88 -
Vicuna v1.3 πŸ¦™ 33B SFT 7.12 88.99
Zephyr-7b-Ξ² πŸͺ 7B DPO 7.34 90.60
WizardLM v1.0 πŸ¦™ 70B SFT 7.71 -
GPT-3.5-turbo - RLHF 7.94 89.37

Specifically, across various categories within the MT-Bench evaluation, Rocket-3B demonstrates impressive performance when compared to larger open models such as Llama2-Chat-7B, Falcon, and Guanaco.

MT-Bench results

MT-Bench detailed score for first and second turn

In MT-Bench, Rocket 🦝 scores 6.99 in the first turn and 6.13 in the second turn, with an average score of 6.56. These scores reflect the model's performance in understanding and generating text during different parts of a conversation.

Model First turn Second turn Average
Rocket 🦝 6.99 6.13 6.56

AlpacaEval detailed scores

In AlpacaEval, Rocket 🦝 achieves a near 80% win rate, coupled with an average response length of 1,242 tokens, indicating its effectiveness in producing detailed responses.

Model Win rate Std error Average length
Rocket 🦝 79.75 1.42 1242

Other benchmarks

Despite its impressive performance on MT-Bench and AlpacaEval benchmarks, the model experiences some challenges when evaluated on other benchmark tests.

Metric Value
Avg. 52.15
ARC (25-shot) 52.82
HellaSwag (10-shot) 73.91
MMLU (5-shot) 61.07
TruthfulQA (0-shot) 57.45
Winogrande (5-shot) 63.22
GSM8K (5-shot) 12.74
DROP (3-shot) 9.66

Intended uses & limitations

Initially, we fine-tuned the model using a dataset created by merging and curating multiple datasets, available on the HuggingFace Hub. This dataset will be released to the public soon. We further enhanced the model's performance using DPO, selecting samples from the openbmb/UltraFeedback and BAAI/JudgeLM-100K datasets. The outcome is a highly effective chat model with a 3 billion parameter scale.

Input Format

The model is trained with the ChatML format:

<|im_start|>system
System message here.<|im_end|>
<|im_start|>user
Your message here!<|im_end|>
<|im_start|>assistant

Here's how you can run the model using πŸ€— Transformers:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model = AutoModelForCausalLM.from_pretrained("pansophic/rocket-3B", trust_remote_code=True, torch_dtype=torch.bfloat16).to("cuda")
tokenizer = AutoTokenizer.from_pretrained("pansophic/rocket-3B", trust_remote_code=True, torch_dtype=torch.bfloat16)
streamer = TextStreamer(tokenizer)

prompt = """<|im_start|>system
{system}<|im_end|>
<|im_start|>user
{user}<|im_end|>
<|im_start|>assistant
"""

system = "You are a helpful assistant."
user = "How are you?"

# Apply the ChatML format
prompt = prompt.format(system=system, user=user)

# Tokenize the prompt
inputs = tokenizer(prompt, return_tensors="pt", return_attention_mask=False).to("cuda")
generated_text = model.generate(**inputs, max_length=3084, top_p=0.95, do_sample=True, temperature=0.7, use_cache=True, streamer=streamer)

# <|im_start|>system
# You are a helpful assistant.<|im_end|>
# <|im_start|>user
# How many helicopters can a human eat in one sitting?<|im_end|>
# <|im_start|>assistant
# Ah, me hearty matey! But yer question be a puzzler! A human cannot eat a helicopter in one sitting, as helicopters are not edible. They be made of metal, plastic, and other materials, not food!<|im_end|>

Bias, Risks, and Limitations

Unlike ChatGPT, which incorporates in-the-loop filtering of responses and is aligned during the RLHF phase for safe completions, our model lacks these features. Consequently, it may generate problematic outputs, particularly when prompted in certain ways.

The pretraining dataset is comprised of a filtered mixture of open-source large-scale datasets available on the HuggingFace Hub: Falcon RefinedWeb extract (Penedo et al., 2023), RedPajama-Data (Together Computer., 2023) and The Pile (Gao et al., 2020) both without the Books3 subset, and StarCoder (Li et al., 2023).

*The model name is inspired by the small but formidable character from 'Guardians of the Galaxy'. Similar to its namesake, this model, with its 3 billion parameters, showcases remarkable efficiency and effectiveness, challenging larger models despite its smaller size."

Model card adapted from Zephyr Beta and Tulu-2-7B