metadata

license: apache-2.0
language:
  - en
library_name: transformers
base_model:
  - Qwen/Qwen2.5-1.5B-Instruct
pipeline_tag: text-generation
model-index:
  - name: Bellatrix-1.5B-xElite
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: IFEval (0-Shot)
          type: wis-k/instruction-following-eval
          split: train
          args:
            num_few_shot: 0
        metrics:
          - type: inst_level_strict_acc and prompt_level_strict_acc
            value: 19.64
            name: averaged accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=prithivMLmods%2FBellatrix-1.5B-xElite
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: BBH (3-Shot)
          type: SaylorTwift/bbh
          split: test
          args:
            num_few_shot: 3
        metrics:
          - type: acc_norm
            value: 9.49
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=prithivMLmods%2FBellatrix-1.5B-xElite
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MATH Lvl 5 (4-Shot)
          type: lighteval/MATH-Hard
          split: test
          args:
            num_few_shot: 4
        metrics:
          - type: exact_match
            value: 12.61
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=prithivMLmods%2FBellatrix-1.5B-xElite
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GPQA (0-shot)
          type: Idavidrein/gpqa
          split: train
          args:
            num_few_shot: 0
        metrics:
          - type: acc_norm
            value: 3.8
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=prithivMLmods%2FBellatrix-1.5B-xElite
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MuSR (0-shot)
          type: TAUR-Lab/MuSR
          args:
            num_few_shot: 0
        metrics:
          - type: acc_norm
            value: 4.44
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=prithivMLmods%2FBellatrix-1.5B-xElite
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU-PRO (5-shot)
          type: TIGER-Lab/MMLU-Pro
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 7.3
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=prithivMLmods%2FBellatrix-1.5B-xElite
          name: Open LLM Leaderboard
tags:
  - qwen
  - qwq

 ____  ____  __    __      __   ____  ____  ____  _  _ 
(  _ \( ___)(  )  (  )    /__\ (_  _)(  _ \(_  _)( \/ )
 ) _ < )__)  )(__  )(__  /(__)\  )(   )   / _)(_  )  ( 
(____/(____)(____)(____)(__)(__)(__) (_)\_)(____)(_/\_)

Bellatrix-1.5B-xElite

Bellatrix-1.5B-xElite is based on a reasoning-based model designed for the QWQ synthetic dataset entries. The pipeline's instruction-tuned, text-only models are optimized for multilingual dialogue use cases, including agentic retrieval and summarization tasks. These models outperform many of the available open-source options. Bellatrix is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions utilize supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF).

Quickstart with Transformers

Here provides a code snippet with apply_chat_template to show you how to load the tokenizer and model and how to generate contents.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/Bellatrix-1.5B-xElite"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "system", "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

Intended Use:

Multilingual Dialogue Systems:
- Designed for conversational AI applications, capable of handling dialogue across multiple languages.
- Useful in customer service, chatbots, and other dialogue-centric use cases.
Reasoning and QWQ Dataset Applications:
- Optimized for tasks requiring logical reasoning and contextual understanding, particularly in synthetic datasets like QWQ.
Agentic Retrieval:
- Supports retrieval-augmented generation tasks, helping systems fetch and synthesize information effectively.
Summarization Tasks:
- Excels in summarizing long or complex text while maintaining coherence and relevance.
Instruction-Following Tasks:
- Can execute tasks based on specific user instructions due to instruction-tuning during training.
Language Generation:
- Suitable for generating coherent and contextually relevant text in various domains and styles.

Limitations:

Synthetic Dataset Bias:
- Optimization for QWQ and similar datasets may make the model less effective on real-world or less structured data.
Data Dependency:
- Performance may degrade on tasks or languages not well-represented in the training dataset.
Computational Requirements:
- The optimized transformer architecture may demand significant computational resources, especially for fine-tuning or large-scale deployments.
Potential Hallucinations:
- Like most auto-regressive models, it may generate plausible-sounding but factually incorrect or nonsensical outputs.
RLHF-Specific Biases:
- Reinforcement Learning with Human Feedback (RLHF) can introduce biases based on the preferences of the annotators involved in the feedback process.
Limited Domain Adaptability:
- While effective in reasoning and dialogue tasks, it may struggle with highly specialized domains or out-of-distribution tasks.
Multilingual Limitations:
- Although optimized for multilingual use, certain low-resource languages may exhibit poorer performance compared to high-resource ones.
Ethical Concerns:
- May inadvertently generate inappropriate or harmful content if safeguards are not applied, particularly in sensitive applications.
Real-Time Usability:
- Latency in inference time could limit its effectiveness in real-time applications or when scaling to large user bases.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here! Summarized results can be found here!

Metric	Value (%)
Average	9.55
IFEval (0-Shot)	19.64
BBH (3-Shot)	9.49
MATH Lvl 5 (4-Shot)	12.61
GPQA (0-shot)	3.80
MuSR (0-shot)	4.44
MMLU-PRO (5-shot)	7.30