migtissera/Tess-v2.5.2-Qwen2-72B

Tess-v2.5.2 (Qwen2-72B)

Update:

This is a fine-tune over the Tess-v2.5, with a changed learning rate and a subset of Tess-v2.5 dataset. The model is completely different to Tess-v2.5.

I was testing a new feature with the Tess-v2.5 dataset. If you had used the model, you might have noticed that the model generations sometimes would end up with a follow-up question. This is intentional, and was created to provide more of a "natural" conversation.

What had happened earlier was that the stop token wasn't getting properly generated, so the model would go on to answer its own question.

This is fixed in Tess-v2.5.2. The model would still ask you follow-up questions, but the stop tokens are getting properly generated. If you'd like to not have the follow-up questions feature, just add the following to your system prompt: "No follow-up questions necessary".

Tess-v2.5.2 (Qwen2-72B)

We've created Tess-v2.5.2, the latest state-of-the-art model in the Tess series of Large Language Models (LLMs). Tess, short for Tesoro (Treasure in Italian), is the flagship LLM series created by Migel Tissera. Tess-v2.5.2 brings significant improvements in reasoning capabilities, coding capabilities and mathematics. It is currently the #1 ranked open weight model when evaluated on MMLU (Massive Multitask Language Understanding). It scores higher than all other open weight models including Qwen2-72B-Instruct, Llama3-70B-Instruct, Mixtral-8x22B-Instruct and DBRX-Instruct. Further, when evaluated on MMLU, Tess-v2.5.2 (Qwen2-72B) model outperforms even the frontier closed models Gemini-1.0-Ultra, Gemini-1.5-Pro, Mistral-Large and Claude-3-Sonnet.

Tess-v2.5.2 (Qwen2-72B) was fine-tuned over the newly released Qwen2-72B base, using the Tess-v2.5 dataset that contain 300K samples spanning multiple topics, including business and management, marketing, history, social sciences, arts, STEM subjects and computer programming. This dataset was synthetically generated using the Sensei framework, using multiple frontier models such as GPT-4-Turbo, Claude-Opus and Mistral-Large.

The compute for this model was generously sponsored by KindoAI.

When evaluated on a subset of AGIEval (Nous), this model compares very well with the godfather GPT-4-0314 model as well.

Training Process

Tess-v2.5.2 model was initiated with the base weights of Qwen2-72B. It was then fine-tuned with the Tess-v2.5 dataset, using Axolotl as the training framework. Most of Tess models follow a common fine-tuning methodology: low learning rates, low number of epochs, and uses very high quality and diverse data. This model was fine-tuned on a 4xA100 VM on Microsoft Azure for 4 days. The model has not been aligned with RLHF or DPO.

The author believes that model's capabilities seem to come primariliy from the pre-training process. This is the foundation for every fine-tune of Tess models, and preserving the entropy of the base models is of paramount to the author.

Sample code to run inference

Note that this model uses ChatML prompt format.

import torch, json
from transformers import AutoModelForCausalLM, AutoTokenizer
from stop_word import StopWordCriteria

model_path = "migtissera/Tess-v2.5.2-Qwen2-72B"
output_file_path = "/home/migel/conversations.jsonl"

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.float16,
    device_map="auto",
    load_in_4bit=False,
    trust_remote_code=True,
)

tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

terminators = [
    tokenizer.convert_tokens_to_ids("<|im_end|>")
]

def generate_text(instruction):
    tokens = 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": 1024,
        "top_k": 50,
    }

    length = len(tokens[0])
    with torch.no_grad():
        rest = 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=tokenizer.eos_token_id,
            eos_token_id=terminators,
        )
    output = rest[0][length:]
    string = tokenizer.decode(output, skip_special_tokens=True)
    return f"{string}"

conversation = f"""<|im_start|>system\nYou are Tesoro, a helful AI assitant. You always provide detailed answers without hesitation.<|im_end|>\n<|im_start|>user\n"""

while True:
    user_input = input("You: ")
    llm_prompt = f"{conversation}{user_input}<|im_end|>\n<|im_start|>assistant\n"
    answer = generate_text(llm_prompt)
    print(answer)
    conversation = f"{llm_prompt}{answer}\n"
    json_data = {"prompt": user_input, "answer": answer}

    with open(output_file_path, "a") as output_file:
        output_file.write(json.dumps(json_data) + "\n")

Join My General AI Discord (NeuroLattice):

https://discord.gg/Hz6GrwGFKD

Limitations & Biases:

While this model aims for accuracy, it can occasionally produce inaccurate or misleading results.

Despite diligent efforts in refining the pretraining data, there remains a possibility for the generation of inappropriate, biased, or offensive content.

Exercise caution and cross-check information when necessary. This is an uncensored model.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	33.28
IFEval (0-Shot)	44.94
BBH (3-Shot)	52.31
MATH Lvl 5 (4-Shot)	27.42
GPQA (0-shot)	13.42
MuSR (0-shot)	10.89
MMLU-PRO (5-shot)	50.68