calibrate_software_engineer.yaml

calibration_set:
  _templates:
    programming_languages: &programming_languages "Solve the following problem using {{ ['Zephyr', 'Prolog', 'Cobol', 'Apex', 'Crystal', 'Fortran', 'Nim', 'Delphi', 'Ada', 'Objective-C', 'VBA', 'Perl', 'Groovy', 'MATLAB', 'Solidity', 'Visual Basic', 'OCaml', 'Erlang', 'Julia', 'Lisp', 'F#', 'Clojure', 'GDScript', 'Scala', 'R', 'Haskell', 'Ruby', 'Elixir', 'Lua', 'Zig', 'Dart', 'Swift', 'Metal', 'PowerShell', 'PHP', 'Kotlin', 'C', 'Java', 'C++', 'C#', 'Bash/Shell', 'Go', 'Rust', 'TypeScript', 'HTML/CSS', 'SQL', 'JavaScript', 'Python', 'Lean', 'Coq', 'Pony', 'D', 'Racket', 'Haxe', 'x86-64 ASM', 'ARM-64 ASM', 'LLVM IR', 'GLSL', 'CUDA', 'Vulkan'][hash(row|string) % 60] }}\n***\n"
    spoken_languages: &spoken_languages "Answer in {{ ['Arabic', 'Chinese', 'French', 'German', 'Greek', 'Hebrew', 'Hindi', 'Japanese', 'Korean', 'Portuguese', 'Russian', 'Spanish', 'Turkish'][hash(row|string) % 13] }}\n***\n"
  max_seq_length: 8192
  shuffle: true
  seed: 42
  datasets:

    # Category Summary (Total: 624 samples)
    # =====================================================
    # General chat (24 samples - 3.85%)
    # Instruction and Reasoning tuning (14 samples - 2.24%)
    # Multilingual (70 samples - 11.22%)
    # Tool use (100 samples - 16.03%)
    # Code / Programming / Software Engineering / Devops (328 samples - 52.56%)
    # Math (12 samples - 1.92%)
    # Sciences (16 samples - 2.56%)
    # Medical (8 samples - 1.28%)
    # Finance (8 samples - 1.28%)
    # Business (16 samples - 2.56%)
    # Humanities and Philosophy (8 samples - 1.28%)
    # Creative Writing, Adventure, Roleplay (13 samples - 2.08%)
    # General Knowledge and Pop Culture (2 samples - 0.32%)
    # Behavioral skills (4 samples - 0.64%)
    # Misc (1 sample - 0.16%)
    # =====================================================

    # Research
    # =====================================================
    # According to this presentation https://minjiazhang.github.io/courses/fall24-resource/slides/awq.pdf
    # AWQ only needs 64 samples to identify salient weights that need to be preserved.
    #
    # This research predates the boom of MoE (Mixture-of-Experts) models
    # and it's safer to assume that 64 samples of a general dataset
    # cannot properly identify salient weights of experts.

    # General chat (24 samples)
    # ---------------------------------------------------------------------------
    - dataset: HuggingFaceH4/ultrachat_200k
      columns: [messages]
      split: train_sft
      formatter: chat_completion
      num_samples: 8
      streaming: true

    - dataset: databricks/databricks-dolly-15k
      split: train
      columns: [instruction, response]
      formatter: prompt_answer
      num_samples: 8

    - dataset: neuralmagic/calibration
      subset: LLM
      split: train
      columns: [messages]
      formatter: chat_completion
      num_samples: 8

    # Instruction and Reasoning tuning (14 samples)
    # ---------------------------------------------------------------------------
    - dataset: HuggingFaceH4/no_robots
      split: train
      columns: [messages]
      formatter: chat_completion
      num_samples: 2

    - dataset: nvidia/HelpSteer
      split: train
      columns: [prompt, response]
      formatter: prompt_answer
      num_samples: 2
      streaming: true

    - dataset: garage-bAInd/Open-Platypus
      split: train
      columns: [instruction, output]
      formatter: prompt_answer
      num_samples: 2

    - dataset: PJMixers/grimulkan_physical-reasoning-ShareGPT
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 4

    - dataset: PJMixers/grimulkan_theory-of-mind-ShareGPT
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 4

    # Multilingual (70 samples)
    # ---------------------------------------------------------------------------
    - dataset: HuggingFaceH4/Multilingual-Thinking
      split: train
      columns: [user]
      formatter: raw_text
      num_samples: 32
      formatter_params:
        prefix: *spoken_languages

    - dataset: ServiceNow-AI/M2Lingual
      subset: full_data
      split: train
      columns: [conversation]
      formatter: chat_completion
      num_samples: 4
      streaming: true

    - dataset: droussis/euroblocks_sft_1sample_per_lang
      split: train
      columns: [conversations]
      formatter: chat_completion
      num_samples: 34

    # Tool use (include commented out ToolAce) (100 samples)
    # ---------------------------------------------------------------------------

    # Fail with minimax!
    # jinja2.exceptions.TemplateError: Message has tool role, but there was no previous assistant message with a tool call!
    # - dataset: Team-ACE/ToolACE
    #   split: train
    #   columns: [system, conversations]
    #   formatter: chat_completion_with_sysprompt
    #   num_samples: 100

    - dataset: interstellarninja/hermes_reasoning_tool_use
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 100
      streaming: true

    # Code / Programming / Software Engineering / Devops (336 samples)
    # ---------------------------------------------------------------------------

    - dataset: deepmind/code_contests
      split: train
      columns: [name]
      formatter: deepmind_code_contests
      num_samples: 50
      streaming: true

    - dataset: dh02391735/stackoverflow-kubernetes-questions
      split: train
      columns: [instruction]
      formatter: raw_text
      num_samples: 8
      streaming: true

    - dataset: diversoailab/humaneval-rust
      split: train
      columns: [prompt]
      formatter: raw_text
      num_samples: 100
      formatter_params: # The dataset actually doesn't hardcode the language
        prefix: *programming_languages

    - dataset: ammarnasr/the-stack-rust-clean
      split: train
      columns: [content]
      formatter: raw_text
      num_samples: 8
      streaming: true
      formatter_params:
        prefix: "Explain this code and comment it for a junior dev.\n***\n"

    - dataset: CSJianYang/CodeArena
      split: test
      columns: [messages]
      formatter: chat_completion
      num_samples: 8

    - dataset: nvidia/OpenCodeInstruct
      split: train
      columns: [input, output]
      formatter: prompt_answer
      num_samples: 8
      streaming: true

    - dataset: nvidia/Llama-Nemotron-Post-Training-Dataset
      split: code
      columns: [input]
      formatter: chat_completion
      num_samples: 8
      streaming: true

    - dataset: nvidia/Nemotron-Competitive-Programming-v1
      split: competitive_coding_cpp_part00
      columns: [messages]
      formatter: chat_completion
      num_samples: 8
      streaming: true

    # The conversations columns has another "conversations" field :/
    # - dataset: sr5434/CodegebraGPT_data
    #   subset: 100k-text
    #   split: train
    #   columns: [conversations]
    #   formatter: sharegpt
    #   num_samples: 8

    - dataset: rombodawg/code_bagel_hermes-2.5
      split: train
      columns: [input, output]
      formatter: prompt_answer
      num_samples: 100
      streaming: true

    - dataset: MathArena/project_euler
      split: train
      columns: [problem]
      formatter: raw_text
      num_samples: 30
      formatter_params:
        prefix: *programming_languages

    # Math (12 samples)
    # ---------------------------------------------------------------------------

    - dataset: nvidia/Llama-Nemotron-Post-Training-Dataset
      split: math
      columns: [input]
      formatter: chat_completion
      num_samples: 4
      streaming: true

    - dataset: nvidia/Nemotron-Math-Proofs-v1
      split: lean
      columns: [formal_statement]
      formatter: raw_text
      num_samples: 4
      streaming: true
      formatter_params:
        prefix: "Can you improve, document and add comment to this Lean proof for a non-mathematician?\n***\n"

    - dataset: nvidia/OpenMathInstruct-2
      split: train
      columns: [problem, generated_solution]
      formatter: prompt_answer
      num_samples: 4
      streaming: true

    # Sciences (16 samples)
    # ---------------------------------------------------------------------------

    - dataset: nvidia/Llama-Nemotron-Post-Training-Dataset
      split: science
      columns: [input]
      formatter: chat_completion
      num_samples: 4
      streaming: true

    - dataset: nvidia/OpenScienceReasoning-2
      split: train
      columns: [input, output]
      formatter: prompt_answer
      num_samples: 8
      streaming: true

    - dataset: MegaScience/MegaScience
      split: train
      columns: [question, answer]
      formatter: prompt_answer
      num_samples: 4
      streaming: true

    # Medical (8 samples)
    # ---------------------------------------------------------------------------

    - dataset: OpenMed/Medical-Reasoning-SFT-GPT-OSS-120B
      split: train
      columns: [messages]
      formatter: chat_completion
      num_samples: 4
      streaming: true

    - dataset: ccdv/pubmed-summarization
      subset: section
      split: train
      columns: [article]
      formatter: raw_text
      num_samples: 4
      streaming: true
      formatter_params:
        prefix: "Summarize this:\n***\n"

    # Finance (8 samples)
    # ---------------------------------------------------------------------------

    - dataset: gbharti/finance-alpaca
      split: train
      columns: [instruction, output]
      formatter: prompt_answer
      num_samples: 4

    - dataset: vladlen32230/summarization-yahoo-stock-finance-article-text
      split: train
      columns: [text]
      formatter: raw_text
      num_samples: 4
      formatter_params:
        prefix: "Summarize this:\n***\n"

    # Business (16 samples)
    # ---------------------------------------------------------------------------

    - dataset: fka/awesome-chatgpt-prompts
      split: train
      columns: [prompt]
      formatter: raw_text
      num_samples: 8

    - dataset: theoldmandthesea/17k_business_book
      split: train
      columns: [question, answer]
      formatter: prompt_answer
      num_samples: 8

    # Humanities and Philosophy (8 samples)
    # ---------------------------------------------------------------------------

    - dataset: ruggsea/stanford-encyclopedia-of-philosophy_instruct
      split: train
      columns: [question, answer]
      formatter: prompt_answer
      num_samples: 2
      streaming: true

    - dataset: mlfoundations-dev/stackexchange_philosophy
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 2

    - dataset: FreedomIntelligence/SocraticChat
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 4
      streaming: true

    # Creative Writing, Adventure, Roleplay (13 samples)
    # ---------------------------------------------------------------------------

    - dataset: Gryphe/Opus-WritingPrompts
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 2

    - dataset: anthracite-org/nopm_claude_writing_fixed
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 2

    - dataset: zerofata/Roleplay-Anime-Characters
      split: train
      columns: [messages]
      formatter: chat_completion
      num_samples: 1

    - dataset: zerofata/Instruct-Anime
      split: train
      columns: [messages]
      formatter: chat_completion
      num_samples: 1

    - dataset: zerofata/Instruct-Anime-CreativeWriting
      split: train
      columns: [messages]
      formatter: chat_completion
      num_samples: 1

    - dataset: sam-paech/gutenberg3-generalfiction-scifi-fantasy-romance-adventure-dpo
      split: train
      columns: [chosen]
      formatter: chat_completion
      num_samples: 2

    - dataset: PocketDoc/Dans-Prosemaxx-Adventure
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 2

    - dataset: anthracite-org/stheno-filtered-v1.1
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 2
      streaming: true

    # General Knowledge and Pop Culture (2 samples)
    # ---------------------------------------------------------------------------

    - dataset: KaraKaraWitch/TvTroper-2025
      split: train
      columns: [article]
      formatter: raw_text
      num_samples: 2
      streaming: true
      formatter_params:
        prefix: "Explain this trope like I'm your grandmother\n***\n"

    # Behavioral skills (4 samples)
    # ---------------------------------------------------------------------------

    - dataset: AquaV/US-Army-Survival-Sharegpt
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 1

    - dataset: AquaV/Interrogation-Sharegpt
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 1

    - dataset: AquaV/Multi-Environment-Operations-Sharegpt
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 1

    - dataset: AquaV/Resistance-Sharegpt
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 1

    # Misc (1 sample)
    # ---------------------------------------------------------------------------

    - dataset: PocketDoc/Dans-Kinomaxx-VanillaBackrooms
      split: train
      columns: [conversations]
      formatter: sharegpt
      num_samples: 1