---
license: cc-by-4.0
extra_gated_prompt: >-
  This dataset contains evaluation traces on the APEX-Agents benchmark, including
  rubric criteria and judge rationales. Like the canonical mercor/apex-agents
  dataset, it is intended exclusively for model evaluation. Any use of this
  dataset for training, fine-tuning, or parameter fitting is forbidden. Crawling
  or scraping the dataset is also forbidden. By requesting access you agree to
  these terms.
task_categories:
  - text-generation
tags:
  - agents
  - apex-agents
  - trajectories
  - reinforcement-learning
pretty_name: ApexAgents Recipe — APEX-Agents 480 Eval Traces
---

# ApexAgents Recipe — APEX-Agents 480 Eval Traces

Full evaluation trial dumps on the **480-task APEX-Agents benchmark** for four model
configurations: Qwen3.5-397B and Qwen3.6-35B, each before and after RL post-training with
the [ApexAgents SkyRL recipe](https://github.com/Mercor-Intelligence/ApexAgents-SkyRL-Recipe).
Each set contains **3 independent evaluation runs** over all 480 tasks (one tarball per run).

Companion dataset: [ApexAgentsRecipe-TBench2_1-EvalTraces](https://huggingface.co/datasets/mercor/ApexAgentsRecipe-TBench2_1-EvalTraces)
(the same four models evaluated on Terminal-Bench 2.1).

## License and intended use

CC-BY-4.0, with the same restriction as the canonical
[mercor/apex-agents](https://huggingface.co/datasets/mercor/apex-agents) benchmark:
these traces (including rubric criteria and judge rationales embedded in grades) are
**for model evaluation only** — using them for training, fine-tuning, or parameter
fitting is forbidden, as is crawling or scraping.

## Sets

| Folder | Model | Checkpoint | Runs |
| --- | --- | --- | --- |
| `Qwen3p5-397B_untrained_traces` | Qwen3.5-397B-A17B | base | run1–3 |
| `Qwen3p5-397B_trained_traces`   | Qwen3.5-397B-A17B | post-trained | run1–3 |
| `Qwen3p6-35B_untrained_traces`  | Qwen3.6-35B-A3B   | base | run1–3 |
| `Qwen3p6-35B_trained_traces`    | Qwen3.6-35B-A3B   | post-trained | pass1–3 |

## Results

MeanReward is the average reward over the 480 tasks (rewards are fractional, in [0, 1]);
Pass@1 is the fraction of trials with a perfect reward of 1.0. Both are the mean ± sample
standard deviation over the 3 runs.

| Set | MeanReward | Pass@1 |
| --- | --- | --- |
| Qwen3.5-397B untrained | 31.29% ± 1.98 | 16.11% ± 1.48 |
| Qwen3.5-397B trained   | 43.18% ± 0.93 | 27.29% ± 1.04 |
| Qwen3.6-35B untrained  | 28.69% ± 1.29 | 13.96% ± 1.46 |
| Qwen3.6-35B trained    | 38.69% ± 0.57 | 22.71% ± 0.83 |

## Harness

All runs use the [Archipelago](https://github.com/Mercor-Intelligence/archipelago)
MCP-based agent inside a Harbor-style trial runner: `max_steps=250`, temperature 1.0,
`top_p=1.0`, `top_k` disabled, 262k-token context. The agent works through MCP tools
(document readers, spreadsheets, chat/email servers, code execution) inside each task's
simulated company world.

## Layout

Each folder holds one gzipped tarball per evaluation run. Inside a tarball, every trial is
a directory named `<task_name>__<trial_id>/`:

```
<task_name>__<trial_id>/
  config.json            trial configuration (task path, agent/model settings)
  result.json            full trial result (config, timing, exception info)
  trial.log              trial runner log
  agent/
    trajectory.json      the agent's message trajectory
    tito_transitions.json  token-in/token-out transitions (training-format rollout)
    start.log, tito_debug.txt
  verifier/
    reward.json          reward plus per-verifier judge grades and rationales
    grade_log.txt, test-stdout.txt, reward.txt
  artifacts/manifest.json
```

## Reproducing the scores

```python
import glob, json, tarfile

for run_tgz in sorted(glob.glob("Qwen3p6-35B_trained_traces/*.tar.gz")):
    rewards = []
    with tarfile.open(run_tgz) as tf:
        for m in tf.getmembers():
            if m.name.endswith("verifier/reward.json"):
                rewards.append(json.load(tf.extractfile(m)).get("reward") or 0.0)
    mean = sum(rewards) / len(rewards)
    pass1 = sum(r == 1.0 for r in rewards) / len(rewards)
    print(f"{run_tgz}: n={len(rewards)} mean_reward={mean:.4f} pass@1={pass1:.4f}")
```