Tess-R1 Limerick (Llama-3.1-70B)
Introduction
Welcome to the Tess-Reasoning-1 (Tess-R1) series of models. Tess-R1 is designed with test-time compute in mind, and has the capabilities to produce a Chain-of-Thought (CoT) reasoning before producing the final output.
The model is trained to first think step-by-step, and contemplate on its answers. It can also write alternatives after contemplating. Once all the steps have been thought through, it writes the final output.
- Step-by-step, Chain-of-Thought thinking process. Uses
<thinking>
</thinking>
tags to indicate when the model is performing CoT. <contemplation>
</contemplation>
tags are used when the model contemplate on its answers.<alternatively>
</alternatively>
tags are used for alternate suggestions.- Finally,
<output>
</output>
tags are used for the final output
Important Note:
In a multi-turn conversation, only the contents between the <output>
</output>
tags (discarding the tags) should be carried forward. Otherwise the model will see out of distribution input data and will fail.
The model was trained mostly with Chain-of-Thought reasoning data, including the XML tags. However, to generalize model generations, some single-turn and multi-turn data without XML tags were also included. Due to this, in some instances the model does not produce XML tags and does not fully utilize test-time compute capabilities. There is two ways to get around this:
- Include a try/catch statement in your inference script, and only pass on the contents between the
<output>
</output>
tags if it's available. - Use the
<thinking>
tag as the seed in the generation, and force the model to produce outputs with XML tags. i.e:f"{conversation}{user_input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n<thinking>"
Prompt Format
The model uses Llama3 prompt format.
System Message
The system message must be the following:
You are Tess-R1, an advanced AI that was created for complex reasoning. Given a user query, you are able to first create a Chain-of-Thought (CoT) reasoning. Once the CoT is devised, you then proceed to first think about how to answer. While doing this, you have the capability to contemplate on the thought, and also provide alternatives. Once the CoT steps have been thought through, you then respond by creating the final output.
Evaluations
Since the model is trained to use test-time-compute, the evalutations were performed by first setting the system message, and then extracting the contents between the <output>
</output>
tags. Only the contents between the tags were then used for the evaluations.
Tess-R1 Limerick | Claude 3.5 Haiku | GPT-4o mini | |
---|---|---|---|
GPQA | 41.5% | 41.6% | 40.2% |
MMLU | 81.6% | - | 82.0% |
MATH | 64.2% | 69.4% | 70.2% |
MMLU-Pro | 65.6% | 65.0% | - |
HumanEval | 61.0% | 88.1% | 87.2% |
The evaluations were performed using a fork of Glaive's simple-evals
codebase. Many thanks to @winglian for performing the evals. The codebase for evaluations can be found here: https://github.com/winglian/simple-evals
Example to run evaluations:
python run_reflection_eval.py tess_r1_70b --evals gpqa mmlu math
The system message have been edited in the sampler to reflect Tess-R1's system prompt.
Inference
I have included a sample Python script below. This script uses a try/catch statement to carry forward the model generations in a multi-turn conversation.
import torch, json
from transformers import AutoModelForCausalLM, AutoTokenizer
import re
class LLM(object):
def __init__(self, model_path):
self.model = AutoModelForCausalLM.from_pretrained(
model_path,
torch_dtype=torch.bfloat16,
device_map="auto",
load_in_4bit=False,
trust_remote_code=False,
)
self.tokenizer = AutoTokenizer.from_pretrained(
model_path, trust_remote_code=False
)
self.terminators = [
self.tokenizer.convert_tokens_to_ids("<|end_of_text|>"),
self.tokenizer.convert_tokens_to_ids("<|eot_id|>"),
]
def generate_text(self, instruction):
tokens = self.tokenizer.encode(instruction)
tokens = torch.LongTensor(tokens).unsqueeze(0)
tokens = tokens.to("cuda")
instance = {
"input_ids": tokens,
"top_p": 1.0,
"temperature": 0.75,
"generate_len": 4096,
"top_k": 50,
}
length = len(tokens[0])
with torch.no_grad():
rest = self.model.generate(
input_ids=tokens,
max_length=length + instance["generate_len"],
use_cache=True,
do_sample=True,
top_p=instance["top_p"],
temperature=instance["temperature"],
top_k=instance["top_k"],
num_return_sequences=1,
pad_token_id=self.tokenizer.eos_token_id,
eos_token_id=self.terminators,
)
output = rest[0][length:]
string = self.tokenizer.decode(output, skip_special_tokens=True)
return f"{string}"
def extract_output(self, text):
pattern = r"<output>(.*?)</output>"
match = re.search(pattern, text, re.DOTALL)
content = match.group(1).strip()
return content
def respond_llama3(self, user_prompt):
conversation = """<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\nYou are Tess-R1, an advanced AI that was created for complex reasoning. Given a user query, you are able to first create a Chain-of-Thought (CoT) reasoning. Once the CoT is devised, you then proceed to first think about how to answer. While doing this, you have the capability to contemplate on the thought, and also provide alternatives. Once the CoT steps have been thought through, you then respond by creating the final output.<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n"""
llm_prompt = f"{conversation}{user_input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
answer = self.generate_text(llm_prompt)
try:
answer_output = self.extract_output(answer)
return answer_output
except:
return answer
model_path = "neurolattice/Tess-R1-Llama-3.1-70B"
llm = LLM(model_path)
conversation = """<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\nYou are Tess-R1, an advanced AI that was created for complex reasoning. Given a user query, you are able to first create a Chain-of-Thought (CoT) reasoning. Once the CoT is devised, you then proceed to first think about how to answer. While doing this, you have the capability to contemplate on the thought, and also provide alternatives. Once the CoT steps have been thought through, you then respond by creating the final output.<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n"""
while True:
user_input = input("You: ")
llm_prompt = f"{conversation}{user_input}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
answer = llm.generate_text(llm_prompt)
print("=" * 132)
print(answer)
try:
answer_output = llm.extract_output(answer)
print("=" * 132)
print(answer_output)
conversation = f"{llm_prompt}{answer_output}<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n"
except:
conversation = f"{llm_prompt}{answer}<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n"
- Downloads last month
- 81