Datasets:

RESMP-DEV
/

mimo-rl-telemetry-archive

license: other
task_categories:
  - time-series-forecasting
tags:
  - mimo
  - reinforcement-learning
  - training-telemetry
  - metrics
  - observability
size_categories:
  - 1GB<n<10GB

MiMo RL telemetry archive

This dataset preserves the public Xiaomi MiMo RL training telemetry feed from https://mimo.xiaomi.com/rl/. It contains aggregate training telemetry, not the Xiaomi RL task environments, model weights, prompts, completions, or individual trajectories.

Public-release scope

data/mimo.sqlite3: consistent standalone SQLite snapshot.
data/raw-responses.tar: 132,263 distinct content-addressed publisher HTTP response bodies under raw/<sha256[:2]>/<sha256>.json.gz.
presentation/: metric catalog, display metadata, notices, overview, format parity inputs, and the three pinned publisher JavaScript files.
observer-source.tar: standard-library collector, metadata parser, notice and overview generators, SQLite schema, tests, and lockfile.
analysis.sql: example read-only queries.
manifest.json: byte size and SHA-256 for every published file.

Internal W&B synchronization state, operational logs, local launchd plists, local paths, virtual environments, caches, live SQLite sidecars, and private acceptance receipts are intentionally excluded.

The publisher page did not state a data license. The archive is therefore marked license: other; source attribution and the exact capture endpoints are retained under Provenance. Do not redistribute this archive under a more permissive license than the upstream publisher permits.

Snapshot

Table	Rows
`benchmark_points`	168
`cycles`	64,807
`events`	83
`fetches`	499,145
`live_reports`	10,000
`metric_points`	122,730
`metric_revisions`	130,930
`notices`	6
`runs`	2
`series_snapshots`	376
`status_samples`	123,537

The two runs are:

pro@1789468339.334 / mimo-v2.6-pro
flash@1789485389.264 / mimo-v2.6-flash

There are 2,077 distinct metric tags covering actor/critic loss, entropy, KL, reward, return, advantage, score, pass rate, rollout counts, penalties, context lengths, sampler counters, trace counters, benchmarks, and status. The 25 source labels are publisher-anonymized identifiers such as code/dataset-obg8; original task identities were not exposed.

Relationship to XiaomiMiMo/MiMo-V2.6-RL-oss

The public Xiaomi dataset is complementary rather than a duplicate:

Property	This archive	Xiaomi RL OSS
Purpose	Aggregate training telemetry	Task environments and verifiers
Primary rows	122,730 metric points	7,684 task rows
Task prompts	No	Yes
Environment assets	No	41,292 repository files
Aggregate training time series	Yes	No
Raw publisher HTTP provenance	Yes	Not applicable

A complete string-leaf comparison found zero exact public prompt matches and zero exact public instance-ID matches in the local raw response bodies.

Usage

import sqlite3

with sqlite3.connect("file:data/mimo.sqlite3?mode=ro", uri=True) as db:
    rows = db.execute(
        """
        SELECT run_key, step, value
        FROM metrics
        WHERE tag = 'dynsam/avg@n'
        ORDER BY run_key, step
        """
    ).fetchall()

Use SQLite read-only connections. metric_points is the latest complete snapshot; metric_revisions preserves historical observations and corrections. All timestamp-like fields are Unix seconds.

Provenance

Source feed: https://mimo.xiaomi.com/rl/
First archived request: 2026-09-16 20:53:37 UTC
Last successful source fetch: 2026-10-09 04:12:09 UTC
The source began returning HTTP 404 on 2026-10-09 04:22:13 UTC, freezing the observed runs at step 30.
The collector retained failed response bodies as well as successful ones. A small number of retained error bodies are therefore not valid JSON.
Cost fields and benchmark scores are publisher-reported values, not independent measurements.