Exgentic/traces-v2

Exgentic Traces v2 — Full Corpus

OpenTelemetry-shaped execution traces for 10,573 agent runs across 6 benchmarks, 5 agent harnesses, and 15+ models. Every trace is one agent solving one benchmark task — chat turns, tool calls, scoring events, costs, timings, and full request/response messages.

This is the full corpus. It includes every span produced during evaluation:

• The agent under test's LLM calls (the model being benchmarked)
• The user-simulator LLM calls in τ²-bench (the customer role-play)
• The judge LLM calls in BrowseCompPlus (the correctness judge)
• Tool invocations (execute_tool spans) and agent root spans (invoke_agent)
• Every retry attempt (both transient failures and the eventual success)

For agent-only chat calls (simulator + judge + non-chat ops + failed retries filtered out), use the companion dataset Exgentic/agent-llm-traces-v2. Its scripts/build_agent_llm_traces.py is the canonical derivation script — re-running it on this corpus reproduces the agent-only dataset bit-for-bit.

---

Schema

from datasets import load_dataset
ds = load_dataset("Exgentic/traces-v2", split="train")
row = ds[0]
row["spans"][0]["attributes"]                      # native dict — direct .get() access
row["spans"][0]["attributes"]["gen_ai.usage.input_tokens"]   # native int (not str)
row["spans"][0]["attributes"]["gen_ai.response.finish_reasons"]  # native list[str]
row["spans"][0]["status"]                          # native dict {"code": int, "message": str}

Per-span fields:

Field	Type	Notes
span_id / trace_id / parent_span_id	string	OTel ids
name / kind	string	OTel span name + kind
start_time / end_time	string	ISO-8601 UTC
status	struct<code: int, message: string>	OTel status
attributes	struct<...180 typed fields...>	Native dict — see typing table below
resource_attributes	struct<...7 typed fields...>	service/SDK/deployment info
events	string	JSON-encoded list (includes gen_ai.evaluation.result for scoring)

Native types on critical fields

The OTel GenAI semantic conventions specify native types for several attributes. This dataset preserves them — no int(attr) or json.loads(attr) cast needed at consumer side:

Attribute	Type
gen_ai.usage.input_tokens	int64
gen_ai.usage.output_tokens	int64
gen_ai.request.max_tokens	int64
gen_ai.request.temperature	double
gen_ai.response.finish_reasons	list<string>
gen_ai.request.stop_sequences	list<string>
gen_ai.input.messages	string (JSON-encoded — multi-part content shape)
gen_ai.output.messages	string (JSON-encoded)
gen_ai.tool.definitions	string (JSON-encoded)

Model name canonicalization

The same underlying deployment previously appeared under up to three different routing-prefix forms (e.g. Azure/gpt-5.2-…, openai/Azure/gpt-5.2-…, gpt-5.2-…). All model names are now canonicalized to a single form by stripping routing prefix segments. This applies to both gen_ai.request.model / gen_ai.response.model in spans and the per-row top-level models list.

Per-row top-level models

Aggregated freshly from each row's chat spans (not from stale ingest metadata). A row's models list contains exactly the set of gen_ai.request.model / gen_ai.response.model values that appear in its chat spans — no empty lists, no orphan entries.

---

Storage layout

Sessions (rows)	10,573
Total spans	626,278 (chat + invoke_agent + execute_tool)
Compressed size	469 MB across 9 shards
Shards	data/train/{0000..0008}.parquet
Compression	zstd level 12, row_group_size=10, data_page_size=4 MB
Source framework	Exgentic
Tracing spec	OTel GenAI semantic conventions

Why these writer parameters

Empirically tuned across schema variants, compression algorithms/levels, row-group sizes, and shard counts:

• struct (not map) for attributes: PyArrow's map_ loads as list[tuple] rather than dict; struct gives v1-compatible direct dict access.
• row_group_size=10: keeps each row group's nested data below PyArrow's chunked-output limit (~2 GB), avoiding ArrowNotImplementedError: Nested data conversions not implemented for chunked array outputs.
• data_page_size=4 MB: ~2× faster cold-load on the wide-attribute struct vs the default 1 MB page granularity, with no size penalty.
• zstd level 12: marginal additional compression over level 22 is not worth the much slower write time; decompression cost is identical.

---

Filtering to agent-only in-place

attrs = span["attributes"]
is_agent_chat = (
    attrs.get("gen_ai.operation.name") == "chat"
    and not attrs.get("exgentic.actor.role")   # drops user_simulator + judge
    and attrs.get("gen_ai.input.messages")     # drops Issue #13 stragglers
)

Or use the pre-filtered companion dataset: Exgentic/agent-llm-traces-v2.

---

Citation

@misc{exgentic-traces-v2,
  title  = {Exgentic Traces v2: OpenTelemetry-Shaped Execution Traces for Agent Benchmarks},
  author = {Exgentic team},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Exgentic/traces-v2}
}