Coding Agent Traces: Token Distribution Analysis for Opus and GPT-oss-120B

Coding agent traces generated using Claude Code for Anthropic's Opus and OpenAI's gpt-oss-120B models. The dataset includes counts for input sequence length (ISL), output sequence length (OSL), and uncached new input (ISL_new). It was created by author netpreme and last updated on June 12, 2026.

Use Cases

• Benchmarking token efficiency of coding agents based on ISL and OSL metrics
• Comparing multi-turn task performance between different LLM architectures like Opus and GPT-oss-120B
• Analyzing the distribution of new versus cached input tokens in agent interactions

Strengths

• Includes traces from two distinct, high-profile LLMs: Anthropic's Opus and OpenAI's gpt-oss-120B
• Provides specific metrics for analysis: ISL, OSL, and ISL_new counts
• Focuses on multi-turn coding agent tasks, a relevant use case for LLM evaluation

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Description metadata is limited; actual data quality requires manual inspection after download

Provenance

Source

huggingface

Collection Method

Traces generated by coding agents using Claude Code, with locally saved files from a harness used for Opus analysis.

Freshness

Last updated 2026-06-12 21:03:39; freshness should be verified