# Mlflow To Atif > Convert bounded live MLflow traces or exported Trace.to_dict() JSON into one canonical ATIF trajectory per trace for Harbor or Eval Author audit coverage. Use for MLflow-to-ATIF conversion, not Intake ingestion. - Skill: `nvidia-nemo/mlflow-to-atif` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add nvidia-nemo/mlflow-to-atif` - Raw SKILL.md: https://api.skillmd.com/api/skills/nvidia-nemo/mlflow-to-atif/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - License: Apache-2.0 - Author: NVIDIA NeMo (https://skillmd.com/u/nvidia-nemo) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/nvidia-nemo/mlflow-to-atif --- # Convert MLflow to ATIF Produce canonical ATIF without routing trace data through Intake. The bundled script writes one owner-private `.atif.json` file per MLflow trace and prints only a content-free summary. ## Protect the trace Treat the source and converted files as restricted data unless the user proves otherwise. Do not print trace payloads, place them in Git, or write them into a public or shared output directory. The script makes its output directory mode `0700` and trajectory files mode `0600`. ## Choose the input For exported data, accept one `Trace.to_dict()` value, a JSON array of those values, or an object shaped as `{"traces": [...]}`: ```bash python scripts/convert_mlflow_to_atif.py \ --input \ --output-dir \ --agent-name \ --agent-version ``` For a live MLflow store, require an explicit experiment and bounded time range. Use a Python environment where MLflow is already available; do not install it or change MLflow authentication on the user's behalf: ```bash python scripts/convert_mlflow_to_atif.py \ --tracking-uri \ --experiment-id \ --since \ --until \ --output-dir \ --agent-name \ --agent-version ``` Live conversion requires HTTPS for remote tracking servers, permits HTTP only for loopback hosts, rejects disabled TLS verification, and does not follow HTTP redirects. It uses the paginated `MlflowClient.search_traces()` API so it also works with MLflow versions that predate the fluent API's `return_type="list"`. The script refuses to overwrite an existing trajectory unless the user asks to replace the same conversion and `--overwrite` is supplied. ## Version and validation Emit ATIF v1.7 and add `--validate-with-harbor` when Harbor is installed. ATIF v1.8 is newer and adds audio content parts. This converter currently projects MLflow text and JSON data only, so v1.8 adds no needed representation. Move the converter and downstream consumers to v1.8 together when they support that version; do not relabel v1.7 output as v1.8. ## Conversion contract - Recover the human instruction from the root span or trace input. Fail instead of inventing an instruction when none exists. - Emit LLM spans as agent steps and preserve model and token metrics when MLflow recorded them. - Emit tool, retriever, embedding, reranker, and guardrail spans as deterministic agent steps with paired tool calls and observations. - Flatten the span tree into timestamp order while retaining native IDs, parent IDs, attributes, events, status, and assessments under namespaced `mlflow` metadata. When an event's parent-span bounds prove that its nominal `time_unix_nano` value is actually in microseconds, normalize that value to nanoseconds and record the correction under `extra.mlflow_to_atif.normalization_codes`. - Reject unresolved parents, cycles, and exported search pages with a non-empty `next_page_token`; collect the complete export before conversion. - Distinguish missing, explicit null, and populated tool outputs in each result's `extra.mlflow.output_state`. - Record every known lossy projection under `extra.mlflow_to_atif.loss_codes`. After conversion, pass one emitted file to the downstream consumer as canonical ATIF. Keep restricted originals separate from restricted converted output.