README.md

metadata

license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - code-generation
  - tool-use
  - agent
  - knapsack
  - navigation
  - rule-diagnosis
  - codeact
  - runtime-semantics
pretty_name: Training Traces
size_categories:
  - 1K<n<10K
configs:
  - config_name: knapsack_persistent
    data_files:
      - split: train
        path: data/knapsack/persistent/traces.jsonl
    default: true
  - config_name: knapsack_stateless
    data_files:
      - split: train
        path: data/knapsack/stateless/traces.jsonl
  - config_name: navigation_persistent
    data_files:
      - split: train
        path: data/navigation/persistent/traces.jsonl
  - config_name: navigation_stateless
    data_files:
      - split: train
        path: data/navigation/stateless/traces.jsonl
  - config_name: rule_diagnosis_persistent
    data_files:
      - split: train
        path: data/rule_diagnosis/persistent/traces.jsonl
  - config_name: rule_diagnosis_stateless
    data_files:
      - split: train
        path: data/rule_diagnosis/stateless/traces.jsonl

Training Traces

Anonymous supplementary release for a double-blind workshop submission. This dataset holds the teacher agent traces used to fine-tune the paper's LoRA adapters (see the sibling cap-sweep-eval-data release and the ten adapter repos alongside this one).

6,000 agent traces total (1,000 per family x runtime combination), produced by an LLM teacher solving each of the paper's three agentic task families -- Opaque Knapsack, navigation, and rule diagnosis -- under two interpreter runtime conditions. Each config is a persistent/stateless pair for one family; all six were used to fine-tune the correspondingly-named LoRA adapters.

Key terms

• Persistent runtime: the Python interpreter keeps all variables alive between agent steps. An agent can write total_weight += w and it persists to the next turn.
• Stateless runtime: the interpreter resets after every step; nothing carries over automatically, so the agent must re-establish any state it needs each turn.

Format

Each line is a {"messages": [...]} chat-format record (system / user / assistant / tool turns) suitable for direct SFT ingestion. Records include the full CodeAct-style system prompt, task prompt, the agent's code blocks and tool outputs (including tool exceptions), through to task completion.

Provenance

Released anonymously alongside a NeurIPS workshop submission for reproducibility review. Non-anonymous release (paper citation, full code) will follow after the review process concludes.