metadata

license: mit
license_link: https://huggingface.co/microsoft/phi-4/resolve/main/LICENSE
language:
  - en
pipeline_tag: text-generation
tags:
  - phi
  - nlp
  - math
  - code
  - chat
  - conversational
  - phi3
inference:
  parameters:
    temperature: 0
widget:
  - messages:
      - role: user
        content: How many R's in strawberry? Think step by step.
library_name: transformers
datasets:
  - amphora/QwQ-LongCoT-130K
base_model:
  - microsoft/phi-4
model-index:
  - name: SuperThoughts-CoT-14B-16k-o1-QwQ
    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: 5.15
            name: averaged accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Pinkstack%2FSuperThoughts-CoT-14B-16k-o1-QwQ
          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: 52.85
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Pinkstack%2FSuperThoughts-CoT-14B-16k-o1-QwQ
          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: 40.79
            name: exact match
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Pinkstack%2FSuperThoughts-CoT-14B-16k-o1-QwQ
          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: 19.02
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Pinkstack%2FSuperThoughts-CoT-14B-16k-o1-QwQ
          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: 21.79
            name: acc_norm
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Pinkstack%2FSuperThoughts-CoT-14B-16k-o1-QwQ
          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: 47.43
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard#/?search=Pinkstack%2FSuperThoughts-CoT-14B-16k-o1-QwQ
          name: Open LLM Leaderboard

Please note, the low IFEVAL results is due to this model always reasoning, instruction following is limited, which caused it to have very low ifeval results, this should not matter for most use cases. gguf/final version: https://huggingface.co/Pinkstack/PARM-V2-phi-4-16k-CoT-o1-gguf

This model can be merged with phi-4 based LLMs!

other gguf version: mradermacher/SuperThoughts-CoT-14B-16k-o1-QwQ-GGUF

Phi-4 Technical Report Phi-4 that has been tuned to be more advanced at reasoning.

Unlike other Parm models we had to optimize our fine tuning process to ensure accuracy while still being able to release this model. Training loss: 0.443800

Beats qwen/qwq at MATH & MuSR & GPQA (MuSR being a reasoning benchmark) Evaluation:

the model uses this prompt format: (modified phi-4 prompt)

{{ if .System }}<|system|>
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|user|>
{{ .Prompt }}<|im_end|>
{{ end }}<|assistant|>{{ .CoT }}<|CoT|>
{{ .Response }}<|FinalAnswer|><|im_end|>

It is recommended to use a system prompt like this one:

You are a helpful ai assistant. Make sure to put your finalanswer at the end.

Open LLM Leaderboard Evaluation Results

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

Metric	Value (%)
Average	31.17
IFEval (0-Shot)	5.15
BBH (3-Shot)	52.85
MATH Lvl 5 (4-Shot)	40.79
GPQA (0-shot)	19.02
MuSR (0-shot)	21.79
MMLU-PRO (5-shot)	47.43

🧀 Examples:

(q4_k_m, 10GB rtx 3080, 64GB memory, running inside of MSTY, all use "You are a friendly ai assistant." as the System prompt.) example 1: example 2: example 3: example 4:

All generated locally and pretty quickly too! 😲 Due to our very limited resources we weren't able to evaluate this model (yet..) if you evaluate it please do let us know!

🧀 Information

⚠️ A low temperature must be used to ensure it won't fail at reasoning. we use 0.3 - 0.8!
⚠️ Due to the current prompt format, it may sometimes put <|FinalAnswer|> without providing a final answer at the end, you can ignore this or modify the prompt format.
this is out flagship model, with top-tier reasoning, rivaling gemini-flash-exp-2.0-thinking and o1 mini. results are overall similar to both of them, we are not comparing to qwq as it has much longer results which waste tokens.

Uploaded model

Developed by: Pinkstack
License: MIT
Finetuned from model : microsoft/phi-4

This phi-4 model was trained with Unsloth and Huggingface's TRL library.