Datasets:
Qwen3.5-4B-UD-Q8_K_XL.gguf on SWE-Explore dev-100
One complete benchmark run: Qwen3.5-4B-UD-Q8_K_XL.gguf localising the files responsible for a bug report, across 100 tasks of the SWE-Explore dev-100 suite. Includes the graded metrics, the full agent trajectories, and pi's raw event logs.
Harness: pi. Read-only toolbox
(read, grep, find, ls), 25-step budget, served by llama.cpp
build b8862-abe84f454.
Headline
| estimator | F1 | n |
|---|---|---|
| reweighted (targets the full 848-task set) | 0.1409 | 100 |
| naive | 0.1540 | 100 |
| naive, parsable tasks only | 0.1588 | 97 |
Use the reweighted mean. dev-100 upsamples the higher-scoring non-Python
strata, so the naive mean is biased high; the weighted estimator
Σwᵢscoreᵢ/Σwᵢ recovers the full-set number (stratum_weight is in the metrics
config).
Read this before quoting a number
65 of 100 tasks exhausted the 25-step budget. Capped tasks average 0.1351 F1; uncapped ones 0.1891. A large part of the recall ceiling is budget, not capability — precision is high, the model runs out of moves. Raising the step limit is the single highest-leverage change to this measurement.
tokens_reasoning is 0, and that is a true statement about the channel. This
server build emits no separate reasoning channel: there are no thinking content
parts and no <think> tags anywhere in the run. The model's analysis is inline in
the assistant text and is already counted in tokens_output. Do not add
anything back — that would double-count.
3 tasks have m_parse_failed = true, and they are real format failures, not
truncation. Each ended with stopReason: stop (never length) and non-empty
content, having emitted a literal <tool_call> XML block as prose instead of the
required answer block. They score 0.0 correctly. Two of them also hit the step cap.
Group by category, not m_source. swe_explore omits source from the
breakdown of a parse-failed task, so those tasks vanish from a by-source mean. The
source_or_category column falls back for you. By category:
| category | mean F1 | n |
|---|---|---|
| multilingual | 0.1453 | 43 |
| pro | 0.2041 | 29 |
| verified | 0.1156 | 28 |
metrics.final_text_capped is truncated at 2000 chars (final_text_is_capped
flags which). The untruncated reply is trajectories.final_text — it matters for
the parse-failure analysis, where one task's real answer is 7046 chars.
The system prompt is not in this dataset. It is unrecoverable from pi's event
stream, which starts at the user turn, and this run was scored without
--capture-messages. trajectories.system_prompt is "" — visibly empty rather
than silently absent.
cost is 0.0 by policy, not because it was free. The harness does not price a
local endpoint; a made-up number would look like a measurement.
Not comparable to
- The SWE-Explore paper's Table 5. No row exists for a model this size, and those figures are n=848 on a different scaffold.
- Any published 100-task row for this benchmark. The old 100-task suite shares only 21 of 100 task ids with today's dev-100 — a 79% different task set, against a per-task F1 SD of ~0.15. A cross-suite delta is a difficulty draw, not a result.
Compare only against another dev-100 run.
Provenance
Reported honestly, because it affects what the artifacts are:
results.jsonwas reconstructed fromresults.partial.jsonlafter the host ran out of memory and killed the run at 99/100. The harness writesresults.jsononly at the end, so the completed work had to be recovered.- 7 tasks were re-run and merged under replace-only-if-improved: 4 had been
turned into cancellation artifacts by a failed stop attempt, 1 timed out
(
astropy-14598, later scoring 0.745), 1 errored, 1 never ran.summary.retry_in_placerecords each decision. - Header fields (
tools,step_limit,served_models) were restored from the verified launch configuration, not read back from the original run — the rebuild could not know them.header.notesays so. - Latency was contended. Another job shared the GPU for part of the run, so
mean_agent_elapsed_s(299.36s) is an upper bound. Scores are unaffected. - 3 tasks needed in-process request retries (
trajectories.auto_retry_attempts); all recovered.
Files
| path | what |
|---|---|
metrics/ |
one row per task — all breakdown metrics, tokens, sampling strata |
trajectories/ |
one row per task — nested messages, untruncated final_text, retry provenance |
runs/ |
one row — the run header and configuration |
raw_logs/ |
pi's event stream per task, streaming deltas removed (2449 MB → 50 MB, lossless) |
results.json, report.md |
the original scored record |
raw_logs/ is enough to re-grade: swe_explore's grader consumes only the
final answer text. The repository corpus is not included (35 GB at full size) and
is not needed unless you want to re-derive the metrics from scratch, in which case
each task id carries its base commit.
Per-message tokens in trajectories use tokens_source: "usage_exact" — pi
reports usage per assistant turn, so nothing is estimated. Non-assistant messages
carry -1/not_reported rather than 0, because a tool result is not free, it is
unattributed.
Totals
Input 3,954,542 · output 335,558 · cache-read 27,045,915 · total 31,336,015 tokens over 2,270 agent steps.
Privacy
Host-identifying fields were rewritten before publication: the serving endpoint's private address, and the absolute log paths (now dataset-relative). Agent-visible content was not rewritten — some task repositories legitimately contain RFC1918 addresses in test fixtures, and altering them would falsify the trajectory. No credentials appear in any artifact; the harness keeps the API key out of the agent's environment by design.
- Downloads last month
- 40