hamishivi/agent-task-openthoughts-agent-rl-5k

OpenThoughts-Agent-RL-5K for tmax

A conversion of all 5,000 tasks from open-thoughts/OpenThoughts-Agent-RL-5K, pinned to revision 409012538183a68a19c5432f88ba0791a824c657, into the same dataset and task-archive format as hamishivi/agent-task-termigen, agent-task-endless-terminals, agent-task-terminal-traj, agent-task-cli-gym, and agent-task-swe-smith.

Format

The train split has messages, ground_truth, dataset, env_config, and source columns. Messages use tmax's Vanillux system and instance templates. ground_truth is the original task ID, dataset is passthrough, env_config.env_name is swerl_vanillux_sandbox, and source is openthoughts_agent_rl_5k.

task-data.tar.gz contains one directory per task, named by its original pymethods2test-XXXX ID. Every source file is preserved byte-for-byte: instructions, metadata, task configuration, Dockerfile, reference solution, and tests. An image.txt is added to each task directory. Reference solutions are retained as upstream artifacts; the tmax sandbox loader does not expose them to the agent. No tasks were filtered or modified.

Container and use

All tasks share the original Dockerfile (python:3.10-slim, working directory /app, pytest installed). The built image is hamishi740/agent-task-openthoughts-agent-rl-5k:409012538183, available for Linux AMD64 and ARM64. This image is set explicitly in every record and image.txt.

from datasets import load_dataset
from huggingface_hub import hf_hub_download

train = load_dataset("hamishivi/agent-task-openthoughts-agent-rl-5k", split="train")
archive = hf_hub_download(
    "hamishivi/agent-task-openthoughts-agent-rl-5k",
    "task-data.tar.gz",
    repo_type="dataset",
)

Extract the archive and configure tmax's sandbox task_data_dir to the directory containing the task directories. Use the same Vanillux training setup as the other agent-task-* datasets. Task-specific verifier timeouts remain in the original task.toml; configure the training harness timeout as appropriate (upstream verifier timeout: 720 seconds).

Validation

All 5,000 records were checked for unique IDs, exact reference schema, Parquet round-trip equality, safe archive paths, and byte-identical preservation of all 35,000 source files. Linux AMD64 container smoke tests covered tasks 0000, 2499, and 4999: empty submissions scored 0 and upstream reference solutions scored 1, with networking disabled. This is format validation and a sampled verifier smoke test, not an evaluation of all 5,000 reference solutions or a full training run. See validation.json and smoke-results.json.

convert.py recreates the conversion using a tmax checkout supplying training/open-instruct/open_instruct/environments/vanillux_prompts.yaml:

uv run --with datasets --with pyyaml python convert.py \
  --repo /path/to/tmax --output ./converted \
  --image hamishi740/agent-task-openthoughts-agent-rl-5k:409012538183

Attribution

Source tasks are released by the OpenThoughts-Agent team under Apache-2.0; see the upstream dataset card for project details and citation. The tmax Vanillux prompt templates are adapted from mini-swe-agent (MIT).

Apptainer images

The current Apptainer pool and unified download manifest are maintained in TMaxxx/agent-task-openthoughts-agent-rl-5k. New SIF uploads go to TMaxxx; earlier images remain available here. Use the downloader and manifest in the linked repository to retrieve all available images with tmax-compatible filenames.