Open-weights baseline (Qwen3.8-27B-FP8, Harbor terminus-2): evidence and card rows; apt retries in Dockerfiles
38559ee verified | license: other | |
| viewer: false | |
| license_name: fasterlabeling-evaluation-license | |
| license_link: LICENSE | |
| pretty_name: "FasterLabeling Showcase: three terminal tasks that frontier models fail" | |
| language: | |
| - en | |
| tags: | |
| - terminal-bench | |
| - harbor | |
| - rl-environments | |
| - agentic-coding | |
| - code | |
| - evaluation | |
| size_categories: | |
| - n<1K | |
| # FasterLabeling showcase: three terminal tasks that frontier models fail | |
| Three Terminal-Bench style tasks in the [Harbor](https://github.com/laude-institute/harbor) task format, each with | |
| a Docker environment, a deterministic grader (a program, not a model judge), a reference solution that scores 1, and | |
| the full record of real frontier-model runs that scored 0. | |
| FasterLabeling builds hard coding and terminal RL environments for frontier labs. This set is the public sample. | |
| Everything in it is ours: seeded hidden items, no personal data, no client work. The only third-party code is the | |
| unmodified open-source release each task reads (pulled at image build time and pinned by hash). | |
| ## The tasks | |
| | task | what the agent must do | hidden items | expert estimate | | |
| |---|---|---|---| | |
| | `never-learned-replies` | A nightly fine-tuning job trains on an endless canary chat. From the sources of TRL, transformers and tokenizers alone, decide which replies some history can never learn. | 310 plans, exact answers | ~3 h | | |
| | `attention-mask-fill-signoff` | For 40 CPU transformer deployments, give the exact interval of additive-mask fill values that are safe, from what the pinned transformers and PyTorch CPU kernels compute in float16 and bfloat16. | 40 deployments, exact intervals | ~4 h | | |
| | `ram-boot-image` | A generator codes a RAM or ROM in many Verilog styles. From the Yosys 0.70 source alone (no synthesis tool, no compiler), predict what four FPGA flows make of it: cell type, cell count, flip-flop bits, and the exact INIT contents of every cell. | 256 queries, exact answers | ~3 h | | |
| Each task is graded all-or-nothing by `tests/test.sh`, which writes `/logs/verifier/reward.txt`. The grader and its | |
| answer key never enter the agent's image. The grader runtime is pinned by SHA-256 (`tests/data/runtime_pins.txt`) | |
| and checked before grading. | |
| ## Results | |
| Every run: the task's own container with the agent added, the instruction file as the only input, no network except | |
| the model's API, 2 hours of agent time, 4 CPUs, 8 GB. Each deliverable was graded twice, in the agent's container and | |
| in a fresh container that received only the deliverable, with identical results. "Wrong" counts the items the grader | |
| rejected; one wrong item is a score of 0. | |
| | task | model | runs | outcome | | |
| |---|---|---|---| | |
| | never-learned-replies | claude-fable-5.1 | 2 | fail, 57 / 310 and 125 / 310 wrong | | |
| | never-learned-replies | claude-opus-5.5 | 3 | fail, 4 / 310, 20 / 310, 12 / 310 wrong | | |
| | never-learned-replies | claude-opus-5.0 | 1 | fail, 16 / 310 wrong | | |
| | attention-mask-fill-signoff | claude-opus-5.5 | 2 | fail, 10 / 40 wrong each | | |
| | attention-mask-fill-signoff | claude-opus-5.0 | 1 | fail, 10 / 40 wrong | | |
| | attention-mask-fill-signoff | muse-wonder-full | 1 | fail, 12 / 40 wrong | | |
| | attention-mask-fill-signoff | muse-wonder | 1 | fail, 10 / 40 wrong | | |
| | ram-boot-image | claude-opus-5.5 | 2 | fail, 38 / 256 and 9 / 256 wrong | | |
| | ram-boot-image | claude-fable-5.1 | 2 | fail, 9 / 256 and 38 / 256 wrong | | |
| | never-learned-replies | qwen3.8-27b-fp8 (open weights) | 1 | fail, 310 / 310 wrong (solution skeleton left unimplemented) | | |
| | attention-mask-fill-signoff | qwen3.8-27b-fp8 (open weights) | 1 | fail, no answer file written | | |
| | ram-boot-image | qwen3.8-27b-fp8 (open weights) | 1 | fail, 256 / 256 wrong (no usable estimator) | | |
| **Open-weights baseline, reproducible by anyone.** The `qwen3.8-27b-fp8` rows come from Qwen3.8-27B-FP8 (Apache 2.0) served | |
| with vLLM 0.22.1 on two 24 GB GPUs, driven by Harbor's own `terminus-2` agent at temperature 0.2 with the same two-hour | |
| limit. All three runs were stopped by the time limit while still working. The full agent transcript of each run is in | |
| `evidence/qwen3.8-27b-fp8/run1/trajectory.json`, so the result can be checked line by line and rerun with the command | |
| below. | |
| The `evidence/` folder of each task holds, per run, the summary, the grader log with the first failures listed | |
| (expected versus given), the reward file, and the model's deliverable exactly as graded. | |
| ## Run it yourself | |
| The same three tasks are published on the Harbor registry as `fasterlabeling/never-learned-replies`, | |
| `fasterlabeling/attention-mask-fill-signoff` and `fasterlabeling/ram-boot-image` | |
| ([hub.harborframework.com](https://hub.harborframework.com/tasks/fasterlabeling/ram-boot-image)), so they can also be | |
| pulled by name. | |
| ```bash | |
| pip install harbor | |
| harbor run -p ./never-learned-replies -a oracle # reference solution, expect reward 1 | |
| harbor run -p ./never-learned-replies -a terminus-2 -m <your-model> | |
| # the open-weights baseline above, against a local vLLM server: | |
| harbor run -p ./never-learned-replies -a terminus-2 -m openai/<served-name> --ak api_base=http://<host>:8000/v1 --ae OPENAI_API_KEY=none | |
| ``` | |
| Any agent that Harbor supports, and any model behind an OpenAI-compatible endpoint, works unchanged. Each task folder | |
| is self-contained: `instruction.md`, `task.toml`, `environment/Dockerfile`, `solution/solve.sh`, `tests/test.sh`. | |
| ## License | |
| Evaluation use is permitted: run any model against these tasks, publish the numbers, cite the set. Use of the tasks, | |
| solutions, graders or transcripts as training data, and redistribution or resale, require written agreement. See | |
| `LICENSE`. | |
| ## Built to order | |
| We build tasks like these to order in the same shape, calibrated to a target pass rate on the buyer's model, with | |
| grader logs from independent frontier runs delivered with every task. Contact: support@fasterlabeling.com, | |
| [fasterlabeling.com](https://fasterlabeling.com). | |