The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.12/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 82, in _split_generators
raise ValueError(
ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 65, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Terminus2 baseline vs Patch-Memory on Terminal-Shift (gpt-5.5)
This dataset contains the raw harbor trial outputs for a head-to-head comparison
of two Terminus2 agent variants on the
wufeiwu/terminal-shift
benchmark, both driven by the same gpt-5.5 model through the Salesforce
Research OpenAI gateway.
terminus2_baselineβ vanilla Terminus2 agent, no cross-trial state.terminus2_pmβ Terminus2 + Patch Memory (harbor_patch_memory.agents.terminus2_pm:Terminus2PatchMemory). Within a single chain (prototype + 4 evolutions/variants) Patch Memory carries a rolling summary of code patches and lessons learned across consecutive trials.
Both runs were executed on the same login node, with the same Docker daemon and
identical task images (alexgshaw/<task>:20251031).
Dataset
- 88 non-singleton chains, 440 tasks total.
- Each chain has a prototype task plus 4 variants (
-EVO-Nor-v-<token>). - Variants are run sequentially within a chain; chains run in parallel.
Files
| File | Contents |
|---|---|
README.md |
this report |
comparison_records.json |
per-task and per-chain JSON dump |
full-terminus2_baseline.tar.gz |
every baseline trial dir (237 MB compressed) β result.json, agent/, verifier/, artifacts/, trial.log, etc. |
full-terminus2_pm.tar.gz |
every pm trial dir (256 MB compressed) β same structure plus a chain-scoped Patch Memory snapshot |
scripts/ |
the launcher, dispatcher, status, and analysis scripts used to produce these runs |
Methodology
- Model:
openai/gpt-5.5via Salesforce gateway (X-Api-Keyheader auth). Both variants used identical sampling / completion settings. - Runner:
harbor trials startper task,setsid-wrapped per chain so chains can be killed cleanly.parallel=2..5chains in flight at any time. - Per-chain ordering: prototype β
-EVO-1..4(alphabetical) β the deterministic order that Patch Memory expects. - Pass criterion:
result.json:verifier_result.rewards.reward >= 1.0. - Coverage caveat: this run was time-bounded (~14 hours wall-clock on a
shared login node); not every chain finished. The fair comparison below
is restricted to the 172 tasks that BOTH variants attempted (i.e. both
produced a
result.json).
Coverage summary
- baseline trials with
result.json: 197 - pm trials with
result.json: 208 - tasks attempted by both variants (intersection): 172
- tasks attempted only by baseline: 25
- tasks attempted only by pm: 36
Aggregate pass rate
Cell format: pass / attempted (rate).
| baseline | pm | |
|---|---|---|
| over all tasks attempted by that variant | 121 / 197 (61.4%) | 130 / 208 (62.5%) |
| over the 172 common tasks (fair, head-to-head) | 108 / 172 (62.8%) | 112 / 172 (65.1%) |
Head-to-head on common tasks
| count | |
|---|---|
| both pass | 93 |
| pm pass, baseline fail (pm wins) | 19 |
| baseline pass, pm fail (baseline wins) | 15 |
| both fail | 45 |
Chain-level comparison
Restricted to the 22 chains where BOTH variants completed all 5 tasks.
| passes / chains | rate | |
|---|---|---|
| baseline all-pass on a chain | 7 / 22 | 31.8% |
| pm all-pass on a chain | 10 / 22 | 45.5% |
β Patch Memory is +13.7 pp at the chain level (where it can leverage prior trials) vs +2.3 pp at the per-task level.
Per-chain pass counts (sorted by pm advantage)
| chain | common tasks | baseline pass | pm pass | delta (pm β base) |
|---|---|---|---|---|
crack-7z-hash |
5/5 | 1 | 5 | +4 |
db-wal-recovery |
5/5 | 3 | 5 | +2 |
cancel-async-tasks |
5/5 | 3 | 5 | +2 |
kv-store-grpc |
5/5 | 4 | 5 | +1 |
git-multibranch |
5/5 | 0 | 1 | +1 |
largest-eigenval |
5/5 | 4 | 5 | +1 |
fix-code-vulnerability |
2/5 | 1 | 2 | +1 |
| ... 16 chains tied at 0 ... | ||||
dna-assembly |
5/5 | 1 | 0 | β1 |
gpt2-codegolf |
5/5 | 2 | 1 | β1 |
fix-git |
5/5 | 5 | 4 | β1 |
extract-elf |
5/5 | 3 | 2 | β1 |
build-pov-ray |
4/5 | 4 | 3 | β1 |
custom-memory-heap-crash |
1/5 | 1 | 0 | β1 |
fix-ocaml-gc |
5/5 | 5 | 3 | β2 |
The full per-chain breakdown is in comparison_records.json under the
chain_stats key.
Highlight: crack-7z-hash
The crispest demonstration of Patch Memory's value: baseline solves only 1/5
trials (the prototype), while pm solves all 5. Inspecting the pm trial dirs you
can see the agent reusing the patch summary written by the prototype trial
(_pm_store/<chain>.jsonl) when attempting each subsequent variant β exactly
what Patch Memory is designed to enable.
Reproducing
# (compute env: docker daemon, harbor CLI, terminal-shift dataset on disk)
cd harbor_patch_memory
pip install -e .
# fill in scripts/terminus2_llm.env with your gateway / model
bash scripts/launch_terminus2_baseline.sh # tmux session: harbor-terminus2-baseline
bash scripts/launch_terminus2_pm.sh # tmux session: harbor-terminus2-pm
# monitor:
python scripts/status.py --variant terminus2_baseline --variant terminus2_pm
# analyse like this dataset:
python scripts/analyze_197.py
The exact scripts used are bundled under scripts/ in this repo.
Notes on the run environment
The runs in this dataset were produced on a shared login node with high
concurrent load (other users' experiments competing for CPU and Docker daemon).
Some chains repeatedly hit the harbor 600s EnvironmentStartTimeoutError
because docker compose up --wait couldn't start fast enough β those tasks
are recorded as failures. A clean re-run on a quieter node would likely raise
both variants' absolute pass rates while leaving the relative comparison
(pm > baseline) intact.
- Downloads last month
- 82