Hugging Face's logo

Datasets:
Andromede-AI
/
messier

format:agent-traces
Modalities:
Document
Tabular
Text
Formats:
parquet
Size:
1M - 10M
ArXiv:
Tags:
agent
agent-traces
ai-agents
agent-evaluation
agent-trajectories
benchmark
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
3

MESSIER v1.0.2

#3
by StefanKrsteski - opened
base: refs/heads/main
←
from: refs/pr/3
Discussion Files changed
+11
-10
Files changed (5) hide show
  1. README.md +4 -3
  2. records.jsonl +2 -2
  3. tasks.jsonl +2 -2
  4. trajectories.jsonl +2 -2
  5. verifiers.jsonl +1 -1
README.md CHANGED
@@ -53,8 +53,8 @@ As agent systems and evaluations evolve, keeping track of their progress is incr
53
  </tr>
54
  <tr>
55
  <td align="center"><strong>72,000</strong><br>Verifiers</td>
56
- <td align="center"><strong>903,202</strong><br>Records</td>
57
- <td align="center"><strong>118,701</strong><br>Trajectories</td>
58
  </tr>
59
  </table>
60
 
@@ -73,9 +73,10 @@ The release is organized into five JSONL tables and a directory of task files:
73
 
74
 
75
  ## News
 
 
76
  - **[26/10/2026]** 🌕 We will present MESSIER at the EMNLP 2026 Main Conference.
77
  - **[09/10/2026]** 🌔 We will present *Predicting Task Difficulty Without Rollouts* as a poster at the COLM 2026 Workshop on Agent Behavior (WAB).
78
- - **[05/10/2026]** 🌖 We release the first public version of MESSIER.
79
  - **[26/09/2026]** 🌓 We release the first public version of MESSIER, including 612 new evaluations across QCircuitBench, DABStep, and MedAgentBench for GPT-5.6 Luna, Qwen3.5-9B, Qwen3.5-4B, Qwen3.5-2B, MiniCPM5-2B, Terminus-Qwen3-8B, OpenThinkerAgent-8B-RL, SERA-8B-GA, and Granite-4.2-8B, using Terminus2 and OpenHands.
80
 
81
  - **[13/09/2026]** 🌒 Updated the data model to include `extra_context`, under which task files are saved. We added original task files from HarveyAI-Lab, GDPval, DABStep, OSWorld, and Toolathlon so the materials available to an agent can be accessed alongside each task.
 
53
  </tr>
54
  <tr>
55
  <td align="center"><strong>72,000</strong><br>Verifiers</td>
56
+ <td align="center"><strong>904,051</strong><br>Records</td>
57
+ <td align="center"><strong>119,482</strong><br>Trajectories</td>
58
  </tr>
59
  </table>
60
 
 
73
 
74
 
75
  ## News
76
+ - **[09/11/2026]** 🌗 This release corrects the treatment of non-scorable attempts in Harbor evaluations.
77
+ - **[05/10/2026]** 🌖 We release the first public version of MESSIER.
78
  - **[26/10/2026]** 🌕 We will present MESSIER at the EMNLP 2026 Main Conference.
79
  - **[09/10/2026]** 🌔 We will present *Predicting Task Difficulty Without Rollouts* as a poster at the COLM 2026 Workshop on Agent Behavior (WAB).
 
80
  - **[26/09/2026]** 🌓 We release the first public version of MESSIER, including 612 new evaluations across QCircuitBench, DABStep, and MedAgentBench for GPT-5.6 Luna, Qwen3.5-9B, Qwen3.5-4B, Qwen3.5-2B, MiniCPM5-2B, Terminus-Qwen3-8B, OpenThinkerAgent-8B-RL, SERA-8B-GA, and Granite-4.2-8B, using Terminus2 and OpenHands.
81
 
82
  - **[13/09/2026]** 🌒 Updated the data model to include `extra_context`, under which task files are saved. We added original task files from HarveyAI-Lab, GDPval, DABStep, OSWorld, and Toolathlon so the materials available to an agent can be accessed alongside each task.
records.jsonl CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:81ccff028973e45a22fae955b2918b02086028299400d8b55927e562dfe57646
3
- size 593534860
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2d9f04103c7726a4cee0e2e5f611364e2a5d7d6290f107d06400f308b05548ea
3
+ size 594575392
tasks.jsonl CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:723728658c05815f4ec24eb05a80140f4ab6f2a2e0accf04bc2aae6f2aa74b80
3
- size 133759514
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:34f5689e290c387985844a58312657967b9015033bf880bd5a5751f4a19b7722
3
+ size 133759499
trajectories.jsonl CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:e05db6cdf00d6634c4a9bbf04a5b52e8af8da7721066fa1ce81f0ec75c037717
3
- size 8529472240
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:09dd14d737c6ebf7639fa2f4720fd027a4235e522b709ec3db120252d95f4e17
3
+ size 8719008595
verifiers.jsonl CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:0e808d76062b53e3a51c74b73d695e35c02514cc069f99aa14089ecf4a9c3da4
3
  size 69162543
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2b4ff628b266f113c7692191412c3feabef6caf1842084ee91c2e16ea6c27bc3
3
  size 69162543