yangjunxiao2021/mle-trajectories

MLE-lite trajectories: sol and Opus5

This run contains 39 Harbor ATIF-v1.7 trajectories from a 21-task MLE-lite evaluation of two native coding agents, plus 42 outcome records and official grading reports. Tasks ran on 2026-09-19–20 using Harbor 0.23.0 with a direct Kubernetes backend.

Model	Native agent	Tasks	Valid submissions / trajectories	Gold / silver / bronze	Any medal
gpt-5.6-sol	Codex 0.155.1	21	21	11 / 3 / 1	15/21
claude-opus-5	Claude Code 2.1.278	21	18	13 / 1 / 0	14/21

Full per-run files are under runs/mle-lite-sol-opus5-20260919.

Files and format

• index.jsonl: one row per model/task, including scores, resources, outcome, trajectory path and checksums. Missing trajectories have a null path.
• trials/<model>/<competition>/agent/trajectory.json: Harbor's original ATIF structure with schema_version, session_id, agent, steps, and final_metrics. The steps contain messages, available reasoning, tool calls, observations and metrics as recorded by the native agent's Harbor adapter.
• trials/<model>/<competition>/verifier/mle-report.json: the official grader's report, including score direction and medal thresholds.
• manifest.json: original and published trajectory SHA256 values, and redaction counts.
• validation.json: all 39 published trajectories passed Harbor 0.23.0's TrajectoryValidator, including its referenced-image checks.

Trajectories preserve the Harbor JSON structure. API keys that appeared in process-list/tool output were replaced with [REDACTED_API_KEY]; affected files are flagged in the index. No task/job configuration files, kubeconfig, authentication files or raw native session logs are included. This release contains trajectories and score metadata, not competition datasets, checkpoints, submission CSVs or separately saved solution files; code that appeared in tool interactions remains in the trajectories.

See Harbor ATIF documentation. ATIF standardizes recorded agent interactions; it does not contain a restorable snapshot of the container filesystem or guarantee every native-client detail was retained.

Outcomes and limitations

• Opus5's NYC taxi, NOMAD and Russian text-normalization tasks exited with code 137 and have no valid submission or saved trajectory. These failures remain in the 21-task denominator. OOM is not established by the retained evidence.
• Opus5's Dog Breed and Plant Pathology tasks reached the eight-hour agent budget but retained valid submissions and trajectories. They are included in the 18 available Opus5 trajectories.
• 11 sol tasks used 16 CPUs / 64 GiB RAM and 10 used 8 CPUs / 32 GiB. All Opus5 tasks used 8 CPUs / 32 GiB. Each requested one RTX 4090. Resource differences are recorded per row.
• The intended agent budget was up to eight hours per task, with grading separate. Agents could finish early; this is not a fixed-compute comparison. agent_seconds is wall time and is not GPU compute time.
• The selected local suite covers 21 of the 22 official MLE-lite tasks; SIIM-ISIC is absent. One selected completed attempt per model/task is reported. Earlier setup failures and cancelled K3 runs are outside this release. Terra and luna did not formally start.
• The official MLE-bench code was pinned to 507f92e1138bb6e40dac5c6ee7a6758e6424bf97. The grader used the pandas compatibility setting pd.options.future.infer_string = False.
• On 18 tasks with valid submissions from both agents, Opus5 had the better reported score in 14, sol in three, and one was tied. Raw scores use different metrics and must not be averaged across competitions.
• Opus5's final trajectory messages describe exploiting repeated clean pages in Denoising and class-count constraints in Leaf. These high scores should be interpreted with the recorded methods and data split; a complete contamination audit or independent solution replay has not been performed. Network access was public.

This dataset is a record of these particular model/agent/resource combinations, not a controlled estimate of standalone model quality. No new license grant over underlying competition material is asserted here.

Load the index and a trajectory

import json
from datasets import load_dataset
from huggingface_hub import hf_hub_download

repo = "yangjunxiao2021/mle-trajectories"
index = load_dataset(repo, "mle-lite-sol-opus5-20260919", split="train")
row = next(r for r in index if r["trajectory_available"])
path = hf_hub_download(repo, row["trajectory_path"], repo_type="dataset")
with open(path) as f:
    trajectory = json.load(f)
print(trajectory["schema_version"], len(trajectory["steps"]))