ahalgh/ale-cli-qwen38-flash

ALE-CLI agent traces, Qwen3.8-Flash-Next

This dataset is the full record of running Qwen3.8-Flash-Next (FP8) in the ALE-Claw harness on the CLI task set of Agents' Last Exam. Each of the 104 tasks was run with 10 seeds, for 1,040 task runs and 1,048 attempts including retries. The 89,795 model calls used 4,703,189,255 prompt tokens, 91.7% of them served from the prefix cache, and generated 102,462,222 tokens.

This repository holds the outcome and call-metadata tables. Model inputs and outputs, trajectories, harness state and agent output files are in ahalgh/ale-cli-qwen38-flash-traces.

units has one row per attempt with its status, score, termination reason and grading record. runs records the serving and harness configuration of each of the 10 jobs. llm_calls has one row per model call with latency, finish reason and token usage, and tokens and tool_calls break those calls down by token type and by tool call. None of these tables contain prompt or completion text. Tables join on seed and run_id together (units, steps, harness_state, agent_outputs); run_id alone repeats across seeds, because seeds launched in the same second share one. They also join on session_id (the comma-separated units.session_ids against llm_calls and llm_payloads) and on call_id (llm_calls, llm_payloads, tokens, tool_calls).

Results

Taking the final attempt for each seed and task, and leaving out 38 runs that hit an infrastructure error and the 2 remaining runs of a task whose LLM-judged grader could not run, the mean score is 54.2% and 29.0% of runs score a full 1.0, over 101 tasks and 1,000 runs. Individual seeds range from 51.1% to 57.3% in mean score and from 25.7% to 31.7% in pass rate.

By ALE difficulty tier, mean score and pass rate are 73.9% and 38.2% on Near-Term (45 tasks), 48.0% and 27.5% on Full-Spectrum (42), and 17.5% and 9.0% on Last-Exam (18). The task list is the CLI list at agents-last-exam commit d9abc07. Upstream has since cut it to 86 tasks, and on the 78 of ours that remain the figures are 46.5% and 19.0%. Both lists are in meta/.

Of the 1,040 final runs, 881 completed, 126 ran out of time and 33 failed. Timeouts keep whatever partial score was recorded. The time limit is each task's own budget: 7,200 seconds for 99 tasks, 14,400 seconds for 3 tasks, 28,800 seconds for 2 tasks.

The figures above come from:

from datasets import load_dataset
u = load_dataset("ahalgh/ale-cli-qwen38-flash", "units", split="train").to_pandas()
u = u[u.is_final & ~u.infra_tainted & u.grader_unavailable.isna()]
u["score"] = u.score.fillna(0.0)
print(u.groupby("task").score.mean().mean())

Collection

The model was served with vLLM across two H200 GPUs, and every call went through a logging proxy that stored the full request and response. The 193 attempts from October 2026 were logged before top-k logprobs and token ids were turned on; the 855 from November have both. In October two jobs on the same node shared a proxy port, so a call's seed is taken from its own run tag rather than from where it was logged. 2,329 calls in 36 sessions belong to attempts that were cut off on 30 October 2026 when the cluster's file quota filled. Those attempts have no row in units, and their seeds and tasks were run again later. Output directories holding more than 5,000 files were replaced with a single manifest file. The sandbox image had 21 paths masked after an earlier run showed that an agent could read a task's reference answer; that run is not included.

Redaction

The operator's username and home, scratch and project paths were replaced with <user>, <host_home>, <host_scratch> and <host_project> throughout. Agents' process listings also captured other users' jobs on the shared compute nodes. Those lines were replaced with <other_user_process>, and other users' paths and names elsewhere with <other_user_path> and <other_user>. Secrets passed as NAME=value or --api-key value, and strings matching common credential formats, were replaced with <redacted:KIND> markers, 10,500 in all. One was a credential exposed on a shared node; most of the rest are a placeholder key and passwords from one task's input files, and false matches inside base64-encoded images, which the marker leaves undecodable. Third-party packages installed inside output directories, and files over 200 MB, are listed with their path, size and sha256 but no content, and omitted_reason says why.

License

The tables are derived from runs of Qwen3.8-Flash-Next, released under the Qwen Community License 1.0, on tasks from Agents' Last Exam, released under CC BY 4.0.