pinchbench-pi-trajectories

Agent trajectories from the pi harness on pinchbench (116 tasks), captured with --capture-messages: one conversation per agent loop, per-message token counts, and the task's graded outcome attached.

844 rows — 812 agent loops plus 32 harness side calls (context-compaction summarisers) — across seven arms, 54 columns.

Arms

Arms are identified by run, a short label; run_note carries the free-text reason the arm exists. Both are per-run, unlike notes, which is the grader's per-task remark — filter is_agent_loop before aggregating either.

run	model	provider	provider_pin	mean
qwen3p6-35b-fireworks-v2	qwen3.6-35b-a3b	Fireworks	—	0.9153
qwen3p6-35b-fireworks	qwen3.6-35b-a3b	Fireworks	—	0.9110
gemma-26b-openrouter-novita-v2	gemma-4-26b-a4b-it	OpenRouter	Novita	0.8314
gemma-26b-openrouter-novita	gemma-4-26b-a4b-it	OpenRouter	Novita	0.8254
gemma-26b-openrouter-unpinned	gemma-4-26b-a4b-it	OpenRouter	—	0.7288
e4b-template-and-path-fixed	gemma-4-E4B-it	HF endpoint (vLLM)	—	0.6863
e4b-baseline-broken-template	gemma-4-E4B-it	HF endpoint (vLLM)	—	0.5856

The -v2 arms are the control, and they are the reason to trust the E4B pair

-v2 means the agent sandbox image, not the model or the provider. v1 hid python/pip/pandas from the agent under the login shell the harness uses; v2 does not. Three models, same change:

model	v1	v2	Δ
qwen3.6-35b-a3b	0.9110	0.9153	+0.004
gemma-4-26b-a4b-it	0.8254	0.8314	+0.006
gemma-4-E4B-it	0.5856	0.6863	+0.101

Both controls sit inside the ±0.18 per-task band — indistinguishable from zero, which is the correct result for a control. v2 is not a generally easier environment; it removed a defect that only bit the model weak enough to hit it. E4B reached for pandas and then thrashed when it was missing; the other two mostly wrote working code first time and never touched the wall.

The two E4B arms are the same model on the same endpoint

They differ only in two serving-stack defects, both found and fixed on 2026-08-08. 0.5856 → 0.6863. Do not average them.

	baseline	fixed
mean score	0.5856	0.6863
trajectories ending in an empty completion	44	0
trajectories hitting No module named 'pandas'	21	0
tool-argument validation errors	370	176
csv_analysis mean	0.4534	0.7401

1. Chat template. vLLM's tool_chat_template_gemma4.jinja emitted no generation prompt when the last message was a tool result and thinking was off, so the prompt ended on a closed turn (<turn|>). <eos> was then the argmax first token (logprob −0.944) and greedy decoding produced an empty completion 5/5. Patched to emit <|turn>model in that case.
2. Sandbox PATH. The harness launches the agent with a login shell (sh -lc); Debian's /etc/profile overwrites PATH, discarding the Dockerfile's ENV PATH=/opt/agent-venv/bin:…. pandas, pip and python were installed but unreachable — csv_analysis lost outright. Fixed via /etc/profile.d plus a build-time smoke test run under sh -lc.

One caveat on the fixed arm: 8 tasks timed out at 16 workers against an endpoint capped at 2 replicas, and were re-run at 4 workers (retried_from); 7 completed in 55–234 s. task_log_nginx_slow_requests timed out again and scores 0.0. suspect_split is true on 4 rows.

⚠️ Two arms carry serving faults — read before comparing

provider names the aggregator, not the backend. OpenRouter fans one model across ~9 serving stacks, and they do not behave identically. Group by (model, provider, provider_pin), never by model alone — the two gemma-4-26b-a4b-it arms differ by 0.097 mean score for serving reasons and would otherwise silently average.

Faults measured on 2026-08-07, all of which present as model failure because the harness sees no usable tool call and treats the turn as finished:

• Unpinned 26b arm — ~1% of calls were served by a vLLM 0.26.0 build (system_fingerprint vllm-0.26.0-845c2a3c / -6ed99b25) whose tool-call parser drops the call: the response arrives empty, or the call text lands in the reasoning channel as call:write{.... 16 trajectories end in an empty final assistant turn. Replaying the same request on a healthy backend returns a normal tool call.
• E4B baseline arm (e4b-baseline-broken-template) — same signature at 4.2% of calls, on a dedicated endpoint running the mutable vllm/vllm-openai:nightly tag (never pinned; drifted ~50 commits between two runs a day apart). 44 trajectories end in an empty final turn. Root-caused to the chat template, not the model — see the E4B section above. Kept as the before-half of that pair; do not read it as a capability measurement.
• The Novita-pinned 26b arm shows none of this: 0 tool-call validation failures across 256 tool results, 0 special-token leaks.

Two providers were tested and rejected, and are not in this dataset: Google (via OpenRouter) truncates tool-call arguments at ~3,725 chars while reporting finish_reason: tool_calls, and a Fireworks deployment leaked special tokens (<|) into argument strings on 100% of edit calls.

reasoning_effort: "none" was requested on every arm but is not honoured uniformly — no OpenRouter provider for gemma-4 declares support for it. The unpinned 26b arm reasoned on 14 calls despite asking for none.

Reading a row

A JSONL capture is not a transcript: chat completions is stateless, so every record repeats the whole history. Rows here are already folded — one row per (model, task, loop), with the conversation in messages and the reply the loop earned appended. is_agent_loop separates the agent's loop from harness side calls; count only agent loops.

Key columns: task, model, run, run_note, provider, provider_pin, score, breakdown_json (per-criterion 0/1), messages, steps, tool_calls, duplicate_call_rate, tokens_*, suspect_split, capture_warnings.

Caveats

• One run per arm, and per-task variance on pinchbench is ±0.18 — a single run cannot separate a few points between two models. This applies to the 0.5856 → 0.6863 E4B pair too: the mechanical counts (44 → 0 empty completions, 21 → 0 pandas failures) are deterministic facts about the trajectories, but the score delta is still one run against one run.
• Provider is a variable, not a constant. The same model scored 0.9115 on Fireworks and 0.725 on OpenRouter in earlier work; this dataset adds a 0.7288 → 0.8254 swing from serving stack alone.
• Check the failure count before reading a mean. Three tasks did not complete in the 26b arms; a low score from a dropped connection and one from a weak model look identical in an average.
• Quantisation is undeclared for several providers and is not recorded here.
• A third defect is present in every E4B arm and is not fixed: vLLM's gemma-4 tool-call parser emits partial results on malformed input rather than rejecting it. An undelimited argument value is truncated at the first }, , or ] — so jq '.items[] | {id, name}' reaches the shell as jq '.items[ — and a call whose name segment lacks { swallows the end marker, producing a bogus function name with empty arguments (8 calls, 1.3%, all in task_log_hdfs_failures).

Built with scripts/build_traj_dataset.py; see docs/trajectory-datasets.md.