OpenEnv Arena: GenAI OTel Trace Lab
Collection
4 items • Updated
Datasets:
Modalities:
Text
Formats:
json
Languages:
English
Size:
< 1K
Tags:
genai-otel-instrument
genai-observability
openenv
reinforcement-learning
opentelemetry
synthetic-data
License:
reference_evidence_path listlengths 2 2 | reference_interpretation stringclasses 8
values | reward dict | spans listlengths 4 4 | split stringclasses 1
value | task stringclasses 8
values | task_id stringclasses 8
values | trace_id stringclasses 8
values | trace_summary stringclasses 8
values |
|---|---|---|---|---|---|---|---|---|
[
"a1",
"a2"
] | zero cost is unpriced and does not mean free inference | {
"evidence_path_ordered_f1": 0.5,
"interpretation_facts": 0.5
} | [
{
"attributes": {},
"duration_ms": 228,
"name": "app.chat",
"parent_span_id": null,
"span_id": "a1",
"status": "OK"
},
{
"attributes": {
"gen_ai.operation.name": "chat",
"gen_ai.request.model": "synthetic-model-a",
"gen_ai.response.model": "synthetic-model-a",
... | train | Determine whether the zero-cost model call is free or unpriced, and cite its causal span path. | genai-01 | synthetic-genai-01 | Trace loaded with related model operations. Inspect the span evidence and parent links before diagnosing. |
[
"b1",
"b2"
] | estimated cost is indicative with no settled billing record, so it is not billable spend | {
"evidence_path_ordered_f1": 0.5,
"interpretation_facts": 0.5
} | [
{
"attributes": {},
"duration_ms": 810,
"name": "agent.run",
"parent_span_id": null,
"span_id": "b1",
"status": "OK"
},
{
"attributes": {
"gen_ai.operation.name": "chat",
"gen_ai.request.model": "local-small-7b",
"gen_ai.response.model": "local-small-7b",
"gen... | train | Classify the reported model-call cost and identify the evidence path. Distinguish the estimate from the billing preview: is this settled billable spend? | genai-02 | synthetic-genai-02 | Trace loaded with related model and accounting operations. Inspect the span evidence and parent links before diagnosing. |
[
"c1",
"c2"
] | requested model is deprecated and needs migration planning | {
"evidence_path_ordered_f1": 0.5,
"interpretation_facts": 0.5
} | [
{
"attributes": {},
"duration_ms": 405,
"name": "assistant.request",
"parent_span_id": null,
"span_id": "c1",
"status": "OK"
},
{
"attributes": {
"gen_ai.operation.name": "chat",
"gen_ai.request.model": "synthetic-legacy-model",
"gen_ai.request.model.deprecated": tr... | train | Identify the release risk in this successful model call and cite the supporting spans. | genai-03 | synthetic-genai-03 | Trace loaded with multiple model and configuration operations. Inspect the span evidence and parent links before diagnosing. |
[
"d1",
"d2"
] | output token count is unavailable on this streamed call, so TPOT is not measured; chunk count does not substitute for output tokens, and the sibling call's TPOT belongs to that separate call | {
"evidence_path_ordered_f1": 0.5,
"interpretation_facts": 0.5
} | [
{
"attributes": {},
"duration_ms": 1240,
"name": "stream.request",
"parent_span_id": null,
"span_id": "d1",
"status": "OK"
},
{
"attributes": {
"gen_ai.operation.name": "chat",
"gen_ai.request.model": "synthetic-stream-model",
"gen_ai.server.time_to_first_token": 0.... | train | For the streamed call with a TPOT-unavailable reason, explain why TPOT is missing. Do not substitute chunk count for output tokens or attribute a sibling call's TPOT to this call; cite the path to the span reporting the unavailable reason. | genai-04 | synthetic-genai-04 | One streamed call records an output-token-count-unavailable reason and chunk count; a sibling call has its own TPOT value. Inspect the span evidence and parent links before diagnosing. |
[
"e1",
"e2"
] | vision capability was recoverably downgraded due to a payload limit | {
"evidence_path_ordered_f1": 0.5,
"interpretation_facts": 0.5
} | [
{
"attributes": {},
"duration_ms": 996,
"name": "agent.turn",
"parent_span_id": null,
"span_id": "e1",
"status": "OK"
},
{
"attributes": {
"gen_ai.operation.name": "chat",
"gen_ai.request.model": "synthetic-multimodal-model",
"gen_ai.response.model": "synthetic-mult... | train | Interpret the successful response's degradation signal and report the affected component and evidence path. | genai-05 | synthetic-genai-05 | Trace loaded with model, modality, and event data. Inspect the span evidence and parent links before diagnosing. |
[
"f1",
"f2"
] | the chat call has a recorded duration and token counts but no TTFT or TPOT measurements; the sibling call records TTFT and TPOT, which apply only to that call | {
"evidence_path_ordered_f1": 0.5,
"interpretation_facts": 0.5
} | [
{
"attributes": {},
"duration_ms": 640,
"name": "chat.endpoint",
"parent_span_id": null,
"span_id": "f1",
"status": "OK"
},
{
"attributes": {
"gen_ai.operation.name": "chat",
"gen_ai.request.model": "synthetic-chat-model",
"gen_ai.system": "openai",
"gen_ai.us... | train | Compare the two model-call spans. For the call with no streaming-latency fields, state which measurements are missing and what is known; cite the causal path to that call. Do not attribute its sibling's TTFT/TPOT values to it. | genai-06 | synthetic-genai-06 | One chat call records duration and token counts only; its sibling records TTFT and TPOT. Inspect the span evidence and parent links before diagnosing. |
[
"g1",
"g2"
] | document payload was omitted after a media upload error | {
"evidence_path_ordered_f1": 0.5,
"interpretation_facts": 0.5
} | [
{
"attributes": {},
"duration_ms": 1440,
"name": "document.answer",
"parent_span_id": null,
"span_id": "g1",
"status": "OK"
},
{
"attributes": {
"gen_ai.media.stripped_reason": "upload_error",
"gen_ai.operation.name": "chat",
"gen_ai.prompt.0.content.0.media_byte_si... | train | Explain why media bytes are absent while identifying the modality and the evidence path. | genai-07 | synthetic-genai-07 | Trace loaded with document and media operations. Inspect the span evidence and parent links before diagnosing. |
[
"h1",
"h2"
] | this is an embedding batch of six inputs rather than chat generation | {
"evidence_path_ordered_f1": 0.5,
"interpretation_facts": 0.5
} | [
{
"attributes": {},
"duration_ms": 188,
"name": "knowledge.lookup",
"parent_span_id": null,
"span_id": "h1",
"status": "OK"
},
{
"attributes": {
"gen_ai.operation.name": "embeddings",
"gen_ai.request.dimensions": 1536,
"gen_ai.request.input_count": 6,
"gen_ai.... | train | Find the embedding operation's batch size and explain whether it is a chat generation call; cite the evidence path. | genai-08 | synthetic-genai-08 | Trace loaded with retrieval and model operations. Inspect the span evidence and parent links before diagnosing. |
GenAI OTel Trace Lab tasks
Eight deterministic synthetic trace-analysis tasks shaped around the public genai-otel-instrument semantic conventions. Each row has a task prompt, four-span synthetic fixture with distractors, canonical interpretation, evidence span path, and reward definition. Reset summaries are neutral and do not state the diagnosis. All IDs, values, span trees, and narratives are synthetic examples based on public instrumentation conventions.
The fixtures teach interpretation of GenAI usage and cost provenance, model deprecation, streaming latency availability, degradation events, multimodal capture metadata, and embedding operations. The cost-provenance task distinguishes an estimate from a billing preview that is not settled. Streaming tasks 04 and 06 explicitly identify the call under analysis; measurements on a sibling span do not establish metrics for that call. Attribute names and conventions follow the public SDK documentation; all trace IDs, values, span trees, and scenario narratives are synthetic.
The environment gives up to 0.5 for curated diagnosis facts and 0.5 for ordered evidence-path F1. This allows supported paraphrases and partial path credit while penalizing extra or reordered spans. Inspection actions are nonterminal; the episode closes with zero reward after eight actions without a diagnosis. Fixtures are mirrored in the environment image; runtime does not download this dataset.
Challenge reference
Built for the OpenEnv Arena challenge Space and its live board. Arena submissions reference a public linux/amd64 OpenEnv server image on Docker Hub or GHCR and include the server schema, task IDs, and example actions. This dataset is a companion collection of synthetic scenarios; the environment image embeds its fixtures and does not download this dataset at runtime.
The published linux/amd64 environment image is kshitij190/trace-path-lab:0.5.2, pinned as docker.io/kshitij190/trace-path-lab@sha256:9dbeab1d60ce2bd3ae1d328ed59457972d7cdaa49b695ca9e6d785109ba648ec. It provides deterministic span-ID and ordering variants for the eight canonical tasks.
The public challenge artifacts are grouped in the OpenEnv Arena: GenAI OTel Trace Lab collection.
Public schema reference: genai-otel-instrument semantic conventions.
Design references
The use of verifiable rewards for GRPO draws on the public RL Wiki knowledge base. Deterministic seed-indexed scenario presentation was informed by the public NoeFlandre OpenEnv Arena Log Triage implementation. The GenAI trace fixtures and checker here are independently authored; no Log Triage code or task data are included.
- Downloads last month
- 26