yizecheng/model-adaptive-tool-necessity

Model-Adaptive Tool Necessity (Raw + Unified JSONL)

This dataset contains the data used in: Model-Adaptive Tool Necessity Reveals the Knowing-Doing Gap in LLM Tool Use.

📝 Paper: https://arxiv.org/abs/2605.14038

💻 Code + original data (GitHub): https://github.com/chengez/Tool-Cognition-Action

Quickstart

from datasets import load_dataset

ds = load_dataset("yizecheng/model-adaptive-tool-necessity")
train = ds["train"]
print(train.column_names)

Main contributions and findings (from the paper)

• Model-adaptive tool necessity: necessity is defined per model, grounded in empirical correctness rather than static annotations.
• Substantial necessity–action mismatch across four models and two domains:
  • Arithmetic: 26.5–54.0% mismatch
  • TruthfulQA: 30.8–41.8% mismatch
• Knowing–doing gap: decodable signals for tool necessity exist in internal representations (“cognition”), but late-layer action representations become nearly orthogonal to these signals (“execution”), leading to failures in translating awareness into direct-answer or tool-call actions.

What’s inside

• train.jsonl: a unified view of the repo’s raw_data/.

  • Each row has:

    • domain (string): dataset domain directory under raw_data/, e.g. math_arithmetic_union or truthfulqa.
    • model (string): model directory name under the domain, e.g. Llama-3.1-8B-Instruct, Qwen3-8B.
    • necessity (nullable string): necessary or unnecessary. This is taken from the filename prefix in raw_data/ (files beginning with necessary_ or unnecessary_).
      • Example source file: raw_data/truthfulqa/Llama-3.1-8B-Instruct/necessary_called.json → necessity = "necessary".
    • action (nullable string): called or notcalled. This is derived from the filename suffix (_called / _Notcalled).
      • Example source file: necessary_Notcalled.json → action = "notcalled".
    • record_type (string): what kind of record this row represents.
      • sample: a sample-level row coming from one of the four per-model split files under raw_data/:
        • necessary_called.json, necessary_Notcalled.json, unnecessary_called.json, unnecessary_Notcalled.json.
        • Example: raw_data/math_arithmetic_union/Qwen3-4B/necessary_called.json → record_type = "sample".
      • output: a model-generation row coming from *_outputs.json files that contain model outputs for a given split and model.
        • Example: necessary_Qwen3-4B_outputs.json or unnecessary_Llama-3.1-8B-Instruct_outputs.json → record_type = "output".
    • source_file / source_path (string): the filename and full path (workspace-relative) of the originating file under raw_data/.
    • id (string): sample identifier shared between sample rows and corresponding output rows.
    • history (string): JSON-encoded value of the original history field (list of message objects). Use json.loads(row['history']) to parse.
    • function (string): JSON-encoded list of function/tool spec objects for the sample. Tool parameter schemas vary across domains (for example some tools use query, others use expression). Encoding as a JSON string ensures a stable top-level schema for HF; parse with json.loads(row['function']).
    • metadata (string): JSON-encoded metadata object from the original files (e.g., correctness counts). Parse with json.loads(row['metadata']).
    • output (string or null): for record_type == 'output', a JSON-encoded object containing the model generation; for sample rows this is null. Parse with json.loads(row['output']) when present.

  • Example filename mappings (raw_data → unified fields):

    • raw_data/<domain>/<model>/necessary_called.json → record_type=sample, necessity=necessary, action=called.
    • raw_data/<domain>/<model>/unnecessary_Qwen3-8B_outputs.json → record_type=output, necessity=unnecessary, action inferred by matching id against the called/notcalled splits.

• raw_data/: original per-model/per-domain files

Example: load the four sample splits for Llama-3.1-8B-Instruct on math_arithmetic_union

from datasets import load_dataset

ds = load_dataset("yizecheng/model-adaptive-tool-necessity")["train"]

rows = ds.filter(
    lambda row: row["record_type"] == "sample"
    and row["domain"] == "math_arithmetic_union"
    and row["model"] == "Llama-3.1-8B-Instruct"
)

called = rows.filter(lambda row: row["necessity"] == "necessary" and row["action"] == "called")
notcalled = rows.filter(lambda row: row["necessity"] == "necessary" and row["action"] == "notcalled")
unnecessary_called = rows.filter(lambda row: row["necessity"] == "unnecessary" and row["action"] == "called")
unnecessary_notcalled = rows.filter(lambda row: row["necessity"] == "unnecessary" and row["action"] == "notcalled")

print("necessary_called", len(called))
print("necessary_notcalled", len(notcalled))
print("unnecessary_called", len(unnecessary_called))
print("unnecessary_notcalled", len(unnecessary_notcalled))