diff --git a/bugsinpy_v2/README.md b/bugsinpy_v2/README.md index b7742b060fb76700b9060fcf785ba2e13566641e..2e60e0664b95058f94a68a2252cd7929ae1060e3 100644 --- a/bugsinpy_v2/README.md +++ b/bugsinpy_v2/README.md @@ -2,39 +2,33 @@ ## Dataset stats -- **v1 tasks evaluated:** 489 -- **Kept:** 65 (solved at least once by either model) -- **Excluded:** 424 (no recorded success) -- **Attempts:** pass@16 with Qwen/Qwen3-30B-A3B-Instruct-2507 and Qwen/Qwen3-30B-A3B (thinking) (15,648 in total) +- **Tasks:** 489 in v1; 65 kept in v2. +- **Teacher evaluation:** Qwen/Qwen3-30B-A3B-Instruct-2507 (thinking off), 16 attempts per task (pass@16); Qwen/Qwen3-30B-A3B (thinking on), 16 attempts per task (pass@16). +- **Coverage:** 489/489 tasks evaluated; 15,648 evaluation attempts. +- **Saved conversations:** 15,648 of 15,648 published attempt records, including successful and unsuccessful attempts. -## Evaluation settings - -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +We keep tasks with at least one verifier-successful attempt from any evaluated model. -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | -| Qwen/Qwen3-30B-A3B | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | on | {"min_p": 0, "temperature": 0.6, "top_k": 20, "top_p": 0.95} | +## Evaluation settings -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 16 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | +| Qwen/Qwen3-30B-A3B | Terminus-2 | On | 16 | 0.6 | 0.95 | 20 | 32,768 | 8,192 | min_p=0 | -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -The table reports settings found in **6 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +**Time limits:** agent 900, 3,600 seconds; verifier 300, 3,600 seconds (sampled configurations; limits vary by task where multiple values are shown). -## Pass@16 filtering +**Output-limit evidence:** The 6 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -| Model | Tasks evaluated | Attempts | Tasks solved at least once | -| --- | ---: | ---: | ---: | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 489 | 7,824 | 56 | -| Qwen/Qwen3-30B-A3B (thinking) | 489 | 7,824 | 47 | -| Either model | 489 | 15,648 | 65 | +## Results -### Successful attempts among kept tasks +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. -| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 tasks | Qwen/Qwen3-30B-A3B (thinking) tasks | +| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 (489 tasks) | Qwen/Qwen3-30B-A3B (489 tasks) | | --- | ---: | ---: | -| 0/16 | 9 | 18 | +| 0/16 | 433 | 442 | | 1/16 | 16 | 17 | | 2/16 | 8 | 7 | | 3/16 | 7 | 4 | @@ -42,6 +36,7 @@ The table reports settings found in **6 sampled saved trial configurations**, no | 5/16 | 6 | 3 | | 6/16 | 5 | 1 | | 7/16 | 1 | 5 | +| 8/16 | 0 | 0 | | 9/16 | 2 | 2 | | 10/16 | 1 | 2 | | 11/16 | 3 | 1 | @@ -49,42 +44,46 @@ The table reports settings found in **6 sampled saved trial configurations**, no | 13/16 | 1 | 0 | | 14/16 | 0 | 1 | | 15/16 | 1 | 0 | +| 16/16 | 0 | 0 | -## Select tasks by successful-attempt count +| Model | Evaluated tasks | Pass@1 | Pass@2 | Pass@4 | Pass@8 | Pass@16 | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | 489 | 3.13% | 4.90% | 7.00% | 9.25% | 11.45% | +| Qwen/Qwen3-30B-A3B | 489 | 2.28% | 3.70% | 5.46% | 7.42% | 9.61% | -You can select tasks by how often they were solved, for example all tasks Qwen/Qwen3-30B-A3B-Instruct-2507 solved exactly once: +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. -```python -import pyarrow.parquet as pq +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)]) -``` +## Files and usage -## Traces +`tasks.parquet` contains 65 kept tasks. -`traces/` holds all 15,648 recorded attempts, successful and failed, from every evaluated task, one row per attempt (Zstandard-compressed Parquet). Columns include `task_id`, `model`, `attempt_id`, `reward`, `verifier_reward`, `retained_v2` and the full `trajectory` and `result` JSON. +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. + +Success-count columns: `instruct_successes`, `thinking_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. ```python -import json -import pyarrow.dataset as ds +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -traces = ds.dataset("traces", format="parquet") -failed = traces.to_table(filter=ds.field("reward") == 0) -conversation = json.loads(failed["trajectory"][0].as_py()) +path = hf_hub_download( + "FWeindel/validated-tasks", + "bugsinpy_v2/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("instruct_successes", "=", 1)]) ``` -Trajectory statistics (plots, command distribution) are in [trace_stats](trace_stats/README.md). +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. -Not yet checked with the repository-leak checks added on 2026-10-17 (leftover git history and file timestamps that could reveal the solution). +## Trace statistics -## Runtime +Of 15,648 recorded attempts, 15,648 have saved conversations and 0 do not. 2,708 attempts have no raw verifier reward. Missing measurements remain missing. -- **Versions:** Apptainer **1.5.4**, Python **3.12.14**, Harbor **0.8.1**. -- **Implementation:** [Marin Harbor `7faf878c14b6`](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) + [our runtime patches](https://github.com/MLI-lab/OT-next-data/tree/main/harbor_patches). +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -| Model | Mean time per attempt | Median time per attempt | Attempts measured | -| --- | ---: | ---: | ---: | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 819.81 s | 591.60 s | 7,824 | -| Qwen/Qwen3-30B-A3B (thinking) | 1,098.50 s | 750.80 s | 7,824 | +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -`patch.py` rebuilds `tasks.parquet` from the current v1 tasks and the recorded results, both pinned to revision `a43f5635`. +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/bugsinpy_v2/plots/readme-pass-at-k.png b/bugsinpy_v2/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..146b7fea50ebcf06a9e36b3d255cf042ae608c0c --- /dev/null +++ b/bugsinpy_v2/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6eaef7e53ccb28a3b7fbcb1a91de8c161b752a049767b6eb153e0c3dabd06cf5 +size 50710 diff --git a/bugsinpy_v2/trace_stats/README.md b/bugsinpy_v2/trace_stats/README.md index 6db6d933eab40624d02f6a6f3452826c0f712d4d..a710e09d01206e8bc892a286add1c0bcb02d89c8 100644 --- a/bugsinpy_v2/trace_stats/README.md +++ b/bugsinpy_v2/trace_stats/README.md @@ -1,49 +1,64 @@ -# Trajectory statistics +# BugsInPy v2 trace statistics -15,648 attempts across 489 tasks of bugsinpy_v2, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3-30B-A3B-Instruct-2507 | Qwen/Qwen3-30B-A3B (thinking) | -| --- | ---: | ---: | -| Recorded attempts | 7,824 | 7,824 | -| Median agent turns | 35 | 32 | -| Median input tokens per attempt | 462,749 | 383,801 | -| Median output tokens per attempt | 21,119 | 35,764 | -| Median agent runtime (seconds) | 346 | 500 | -| Median verifier runtime (seconds) | 18 | 17 | +Of 15,648 recorded attempts, 15,648 have saved conversations and 0 do not. 2,708 attempts have no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -![Trace measurements](plots/all-models-trace-distributions.png) +Token totals sum calls across an attempt and are not the context window. -![Trace outcomes](plots/all-models-trace-outcomes.png) +Measurements cover 489 tasks, including kept and unselected tasks, with each attempt counted once. -![Most frequent commands](plots/all-models-trace-commands.png) +## Evaluation settings -Commands are counted as submitted, not as successful executions. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 16 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | +| Qwen/Qwen3-30B-A3B | Terminus-2 | On | 16 | 0.6 | 0.95 | 20 | 32,768 | 8,192 | min_p=0 | -Per-model plots: [Qwen/Qwen3-30B-A3B-Instruct-2507 commands](plots/qwen3-30b-instruct-2507-trace-commands.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 distributions](plots/qwen3-30b-instruct-2507-trace-distributions.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 outcomes](plots/qwen3-30b-instruct-2507-trace-outcomes.png), [Qwen/Qwen3-30B-A3B (thinking) commands](plots/qwen3-30b-thinking-trace-commands.png), [Qwen/Qwen3-30B-A3B (thinking) distributions](plots/qwen3-30b-thinking-trace-distributions.png), [Qwen/Qwen3-30B-A3B (thinking) outcomes](plots/qwen3-30b-thinking-trace-outcomes.png). +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Data +**Time limits:** agent 900, 3,600 seconds; verifier 300, 3,600 seconds (sampled configurations; limits vary by task where multiple values are shown). -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +**Output-limit evidence:** The 6 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -Attempts without a saved conversation: 0; without a verifier reward: 2,708. Their missing measurements stay empty. +## Measurements -## Evaluation settings +### Qwen/Qwen3-30B-A3B-Instruct-2507 + +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 7,824 | 35 | 142 | +| Input tokens per attempt | 7,824 | 462,749 | 2,324,016 | +| Output tokens per attempt | 7,824 | 21,119 | 107,545 | +| Largest recorded step, prompt plus output tokens | 7,824 | 24,745 | 37,862 | +| Agent runtime (seconds) | 7,824 | 345.77 | 1,414.77 | +| Verifier runtime (seconds) | 7,733 | 18.48 | 249.13 | +| Whole-attempt runtime (seconds) | 7,824 | 591.56 | 1,882.49 | -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +### Qwen/Qwen3-30B-A3B -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | -| Qwen/Qwen3-30B-A3B | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | on | {"min_p": 0, "temperature": 0.6, "top_k": 20, "top_p": 0.95} | +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 7,824 | 32 | 235 | +| Input tokens per attempt | 7,824 | 383,801 | 3,849,233 | +| Output tokens per attempt | 7,824 | 35,764 | 287,655 | +| Largest recorded step, prompt plus output tokens | 7,823 | 22,816 | 25,889 | +| Agent runtime (seconds) | 7,824 | 500.16 | 3,600 | +| Verifier runtime (seconds) | 7,533 | 17.03 | 248.72 | +| Whole-attempt runtime (seconds) | 7,824 | 750.63 | 4,033.94 | -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -The table reports settings found in **6 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](../metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +## Plots + +![Trace measurements](plots/all-models-trace-distributions.png) + +![Trace outcomes](plots/all-models-trace-outcomes.png) + +![Most frequent commands](plots/all-models-trace-commands.png) diff --git a/bugsinpy_v3/README.md b/bugsinpy_v3/README.md index 30c651ab0abe720ac74d77fecaaa80cdb7556f5c..2ed2a6b0662bda325083afd373e97201e5998c75 100644 --- a/bugsinpy_v3/README.md +++ b/bugsinpy_v3/README.md @@ -2,106 +2,75 @@ ## Dataset stats -- **Source tasks:** 489 -- **Kept in v3:** 406 newly solved tasks -- **Available separately in v2:** 65 -- **Combined v2 + v3:** 471 unique tasks -- **Excluded:** 18 (no recorded success) -- **New evaluation:** 424 previously excluded tasks, 2 attempts each, 848 attempts -- **New saved trajectories:** 848, successful and failed +- **Tasks:** 424 had no successful attempt in v2; all 424 evaluated for v3; 406 newly kept; 18 still without success. +- **Teacher evaluation:** Qwen/Qwen3.8-27B (thinking on), 2 attempts per task (pass@2). +- **Coverage:** 424 tasks evaluated; 848 evaluation attempts. +- **Saved conversations:** 848 of 848 published attempt records, including successful and unsuccessful attempts. -BugsInPy contains Python bug-fixing tasks from real open-source projects. Each task provides a buggy project checkout and a verifier for the target bug. `tasks.parquet` contains only newly solved tasks with their evaluated payloads. +We keep tasks with at least one verifier-successful attempt from the new model. `tasks.parquet` contains only newly kept tasks; tasks already kept in v2 remain separate. ## Evaluation settings -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | On | 2 | 1.0 | 0.95 | 20 | 32,768 | 8,192 | min_p=0; presence_penalty=0; repetition_penalty=1 | -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3.8-27B | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | on | {"min_p": 0.0, "presence_penalty": 0.0, "repetition_penalty": 1.0, "temperature": 1.0, "top_k": 20, "top_p": 0.95} | +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +**Time limits:** agent 900 seconds; verifier 300 seconds (sampled configurations; limits vary by task where multiple values are shown). -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +**Output-limit evidence:** The 3 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -The table reports settings found in **3 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +## Results -## Evaluation results +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. Only the new v3 evaluation is included. -| Model | Tasks evaluated | Attempts | Tasks solved at least once | Pass@1 | Pass@k | -| --- | ---: | ---: | ---: | ---: | ---: | -| Qwen/Qwen3.8-27B (thinking) | 424 | 848 | 406 | 90.92% | 95.75% (k=2) | - -Qwen/Qwen3.8-27B solved 406 of the 424 previously excluded tasks. Combining these 406 tasks with the 65 tasks in [v2](../bugsinpy_v2/README.md) gives 471 unique tasks. These results cover only the new evaluation. - -Qwen/Qwen3.8-27B used Terminus-2, thinking enabled, temperature 1.0, top-p 0.95, top-k 20, a 32,768-token context, and an intended limit of 8,192 output tokens per model call (see evaluation settings). Each task kept its existing agent and verifier time limits. The task source is pinned to revision `3abfc632`; complete settings are in [evaluation setup](metadata/evaluation-setup.json). - -### Successful attempts in the new evaluation - -| Successful attempts | Tasks | +| Successful attempts | Qwen/Qwen3.8-27B (424 tasks) | | --- | ---: | | 0/2 | 18 | | 1/2 | 41 | | 2/2 | 365 | -![Qwen/Qwen3.8-27B success distribution](trace_stats/plots/qwen38-success-counts.png) - -## Select tasks by successful-attempt count +| Model | Evaluated tasks | Pass@1 | Pass@2 | +| --- | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | 424 | 90.92% | 95.75% | -The Parquet preserves `instruct_successes`, `thinking_successes`, and `attempts_per_model` from v2, and adds `qwen38_successes`, `qwen38_attempts`, and `kept_from_v2`. Tasks already in v2 are not duplicated in this file. +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. -```python -import pyarrow.parquet as pq +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -# Newly evaluated tasks solved exactly once. -subset = pq.read_table("tasks.parquet", filters=[("qwen38_successes", "=", 1)]) -``` +## Files and usage -[Selection metadata](metadata/selection.json) records counts and membership for all 489 tasks, including the 18 excluded tasks. +`tasks.parquet` contains 406 newly kept tasks only. Tasks already kept in v2 are not duplicated. -## Traces +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. -`traces/` includes all 848 new Qwen/Qwen3.8-27B attempts, successful and failed, one row per attempt in Zstandard-compressed Parquet. Earlier traces remain in v2. Columns include `task_id`, `model`, `attempt_id`, `reward`, `verifier_reward`, `retained_v2`, `retained_v3`, and full `trajectory` and `result` JSON. +Success-count columns: `instruct_successes`, `thinking_successes`, `qwen38_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. ```python -import json -import pyarrow.dataset as ds +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -traces = ds.dataset("traces", format="parquet", exclude_invalid_files=True) -new = traces.to_table(filter=ds.field("model") == "qwen3.8-27b") -conversation = json.loads(new["trajectory"][0].as_py()) +path = hf_hub_download( + "FWeindel/validated-tasks", + "bugsinpy_v3/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("qwen38_successes", "=", 1)]) ``` -All 848 new attempts have saved conversations. The new run recorded 232 agent timeouts, 2 verifier runtime errors, 1 verifier timeout, and 1 repeated-output-truncation error. A task can still receive a successful verifier grade after its agent reaches the time limit. Missing verifier rewards remain null and count as unsolved for pass@k. Traces are included without quality filtering; a saved conversation is not necessarily a successful training example. - -Trajectory statistics, per-task and per-attempt measurements, outcome plots, and command distributions are in [trace_stats](trace_stats/README.md). These are generated with the repository's `validation.publishing.trace_stats` script. - -Not yet checked with the repository-leak checks added on 2026-10-17 (leftover git history and file timestamps that could reveal the solution). - -## Runtime - -- **Versions:** Apptainer **1.5.4**, Python **3.12.14**, Harbor **0.8.1**. -- **Implementation:** [Marin Harbor `7faf878c14b6`](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) and [runtime patches](https://github.com/MLI-lab/OT-next-data/tree/main/harbor_patches). -- **New evaluation resources:** one H200 per shard, 32 allocated CPUs, 4 concurrent trials per GPU, up to 8 simultaneous shards. - -| Model | Mean time per attempt | Median time per attempt | Attempts measured | -| --- | ---: | ---: | ---: | -| Qwen/Qwen3.8-27B | 1,172.06 s | 935.76 s | 848 | - -`patch.py` rebuilds the incremental task file from the pinned pre-migration release and the published layout manifest. - -## Incremental release - -V3 contains only tasks newly solved after v2. It has no task-ID overlap with [v2](../bugsinpy_v2/README.md). This is a harder subset relative to the earlier evaluation, not a claim of intrinsic difficulty. New-evaluation traces include both solved and unsolved tasks so the reported pass rates remain reproducible. +Combine the 65 tasks in `bugsinpy_v2/tasks.parquet` with these 406 newly kept tasks for 471 unique tasks. Allow for different columns and retain each source file's image manifest when staging images. ```python -from huggingface_hub import hf_hub_download import pyarrow as pa -import pyarrow.parquet as pq def load(version): - path = hf_hub_download("FWeindel/validated-tasks", - f"bugsinpy_{version}/tasks.parquet", repo_type="dataset") + path = hf_hub_download( + "FWeindel/validated-tasks", + f"bugsinpy_{version}/tasks.parquet", + repo_type="dataset", + ) return pq.read_table(path) v2, v3 = load("v2"), load("v3") @@ -109,4 +78,14 @@ assert set(v2["path"].to_pylist()).isdisjoint(v3["path"].to_pylist()) combined = pa.concat_tables([v2, v3], promote_options="permissive") ``` -The example combines task rows; each source file retains its own image manifest. Use those source manifests when staging images. [Layout and provenance](metadata/layout.json) records the pinned baseline and task counts. Earlier cumulative v3 releases remain accessible through their HF commit history. +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. + +## Trace statistics + +Of 848 recorded attempts, 848 have saved conversations and 0 do not. 3 attempts have no raw verifier reward. Missing measurements remain missing. + +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. + +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. Historical attempts remain in v2. + +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/bugsinpy_v3/plots/readme-pass-at-k.png b/bugsinpy_v3/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..7e3d38c7d923824fdc4eccec96a5cfd47abd940f --- /dev/null +++ b/bugsinpy_v3/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95c2c8607857856728faa40f92ad01d1a5674808ef593524726ff31726ba18f2 +size 35889 diff --git a/bugsinpy_v3/trace_stats/README.md b/bugsinpy_v3/trace_stats/README.md index 8c7212993690f572f4dbaba08bed0b54df676d74..f0ad06c48a60b194259a392943e0b4be1ac6a6fa 100644 --- a/bugsinpy_v3/trace_stats/README.md +++ b/bugsinpy_v3/trace_stats/README.md @@ -1,52 +1,55 @@ -# Trajectory statistics +# BugsInPy v3 trace statistics -848 attempts across 424 tasks of bugsinpy_v3, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3.8-27B | -| --- | ---: | -| Recorded attempts | 848 | -| Median agent turns | 21 | -| Median input tokens per attempt | 247,841 | -| Median output tokens per attempt | 26,467 | -| Median agent runtime (seconds) | 761 | -| Median verifier runtime (seconds) | 34 | +Of 848 recorded attempts, 848 have saved conversations and 0 do not. 3 attempts have no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. Historical attempts remain in v2. -![Trace measurements](plots/all-models-trace-distributions.png) +Token totals sum calls across an attempt and are not the context window. -![Trace outcomes](plots/all-models-trace-outcomes.png) +Measurements cover 424 tasks, including kept and unselected tasks, with each attempt counted once. -![Most frequent commands](plots/all-models-trace-commands.png) +## Evaluation settings -Commands are counted as submitted, not as successful executions. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | On | 2 | 1.0 | 0.95 | 20 | 32,768 | 8,192 | min_p=0; presence_penalty=0; repetition_penalty=1 | -Per-model plots: [Qwen/Qwen3.8-27B commands](plots/qwen3.8-27b-trace-commands.png), [Qwen/Qwen3.8-27B distributions](plots/qwen3.8-27b-trace-distributions.png), [Qwen/Qwen3.8-27B outcomes](plots/qwen3.8-27b-trace-outcomes.png). +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Data +**Time limits:** agent 900 seconds; verifier 300 seconds (sampled configurations; limits vary by task where multiple values are shown). -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +**Output-limit evidence:** The 3 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -Attempts without a saved conversation: 0; without a verifier reward: 3. Their missing measurements stay empty. +## Measurements +### Qwen/Qwen3.8-27B -Evaluation budgets: see the [dataset evaluation setup](../README.md). Per-call context and output-limit enforcement are not established by token-usage statistics. +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 848 | 21 | 46 | +| Input tokens per attempt | 848 | 247,841 | 706,385 | +| Output tokens per attempt | 848 | 26,467 | 86,804 | +| Largest recorded step, prompt plus output tokens | 848 | 23,226 | 31,148 | +| Agent runtime (seconds) | 848 | 761.22 | 2,320.55 | +| Verifier runtime (seconds) | 848 | 34.42 | 255.65 | +| Whole-attempt runtime (seconds) | 848 | 935.76 | 2,818.4 | +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -## Token usage and summarization +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -All published traces are included, covering kept and excluded tasks and every verifier outcome. One trace means one task attempt. +## Plots -![Output, summarization, and context distributions](plots/all-models-tokens-by-outcome.png) +![Trace measurements](plots/all-models-trace-distributions.png) + +![Trace outcomes](plots/all-models-trace-outcomes.png) -- **Output per recorded agent step:** successful, graded unsuccessful, and ungraded attempts shown separately. Each step contributes once. -- **Summarizations per trace:** one combined distribution across all traces with a recorded count, including zero. -- **Peak context per trace:** the largest recorded step prompt + output count in each trace, separated into traces with and without summarization. Unknown summary counts are kept separate when present. Each trace contributes once, regardless of its length or outcome. +![Most frequent commands](plots/all-models-trace-commands.png) -[Detailed statistics](data/tokens-by-outcome.json) retain sample counts and missing measurements. Missing summary counts are not zero. ATIF steps can aggregate multiple requests, so recorded context does not establish an exact per-call context limit. +![Output, summarization and recorded context](plots/all-models-tokens-by-outcome.png) -This folder reports the new evaluation only. Task membership is incremental: tasks already in v2 are not duplicated in v3. Historical conversations remain in v2. See the [release overview](../README.md) for counts and combination instructions. +Output per recorded agent step is separated by verifier outcome. Summarizations are counted per trace, including zero when recorded. Peak recorded context is shown per trace with and without summarization; unknown counts stay separate. diff --git a/calibforge_v2/README.md b/calibforge_v2/README.md index 7ea6ed1e6c49548f3af5221fb39b2273ea1139b4..5c6b788aacdb3702864760e5ba67d73783226618 100644 --- a/calibforge_v2/README.md +++ b/calibforge_v2/README.md @@ -2,58 +2,32 @@ ## Dataset stats -- **v1 tasks evaluated:** 5,317 -- **Kept:** 1,512 (solved at least once) -- **Excluded:** 3,805 (never solved) -- **Attempts:** pass@8 with Qwen/Qwen3-30B-A3B-Instruct-2507 (42,536 in total) +- **Tasks:** 5,317 in v1; 1,512 kept in v2. +- **Teacher evaluation:** Qwen/Qwen3-30B-A3B-Instruct-2507 (thinking off), 8 attempts per task (pass@8). +- **Coverage:** 5,317/5,317 tasks evaluated; 42,536 evaluation attempts. +- **Saved conversations:** 42,536 of 42,536 published attempt records, including successful and unsuccessful attempts. -## Evaluation settings - -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. - -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | - -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. - -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. - -The table reports settings found in **2 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. - -## Optional subsets for a 25K-context training comparison - -The main `tasks.parquet` remains unchanged. These additional files select existing v2 task rows by their historical successful traces; they are not results of a new capped evaluation. - -| File | Tasks | Selection | -| --- | ---: | --- | -| [tasks_solved_ctx25000_out8192_no_summary.parquet](tasks_solved_ctx25000_out8192_no_summary.parquet) | 1,258 | At least one verifier-successful attempt with complete step metrics, every recorded agent step output **<8,192**, prompt plus output **≤25,000**, and explicitly **zero summaries**. Recommended starting subset. | - -25K here means **25,000 total prompt-plus-output tokens per recorded step**, not 25,600 and not a total rollout-generation budget. ATIF step metrics can aggregate calls/retries. This is a conservative observed-trace filter, not proof that all model requests or future RL rollouts will fit. A successful short attempt is sufficient to keep a task; its other attempts can be longer or unsuccessful. Missing usage and unknown summary counts do not qualify for the recommended subset. - -Across **4,555 successful attempts**, median peak recorded context is **9,245 tokens** (90th percentile **25,126**); median peak step output is **1,213 tokens**. Median summed step output is **3,856 tokens** (90th percentile **27,977**), which is a different measurement from context. **3,606 successes used no summarization** and **949 used summaries**. The strict subset has **3,606 qualifying successful attempts**. +We keep tasks with at least one verifier-successful attempt from any evaluated model. -Task payloads and existing result columns are preserved exactly from the published v2 Parquet. Historical evaluated payload hashes differ from the published payloads, as described in [v2 selection provenance](metadata/selection.json); this release does not establish a new evaluation of those published bytes. Do not interpret success-selected subset rates as unbiased pass@k under enforced 25K/8K limits. - -[Filter definition and statistics](metadata/token-budget-filter/summary.json), [per-task witness attempt IDs and success counts](metadata/token-budget-filter/per-task.parquet), [all-attempt measurements](metadata/token-budget-filter/per-attempt.parquet), and [reproducible selector](metadata/token-budget-filter/selector.py) preserve the selection evidence. The original model's success counts can guide a curriculum but do not establish difficulty for a different RL checkpoint. - -![Recorded trace-length distributions](metadata/token-budget-filter/trace-lengths.png) +## Evaluation settings +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 8 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | -## Pass@8 filtering +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -| Model | Tasks evaluated | Attempts | Tasks solved at least once | -| --- | ---: | ---: | ---: | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 5,317 | 42,536 | 1,512 | +**Time limits:** agent 900, 1,200 seconds; verifier 900 seconds (sampled configurations; limits vary by task where multiple values are shown). -All 5,317 tasks have been evaluated with eight attempts each. +**Output-limit evidence:** The 2 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -14 kept tasks were evaluated on an earlier version of the task; tasks.parquet has the current version. +## Results -### Successful attempts among kept tasks +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. -| Successful attempts | Tasks | +| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 (5,317 tasks) | | --- | ---: | +| 0/8 | 3,805 | | 1/8 | 530 | | 2/8 | 288 | | 3/8 | 207 | @@ -63,40 +37,47 @@ All 5,317 tasks have been evaluated with eight attempts each. | 7/8 | 87 | | 8/8 | 92 | -## Select tasks by successful-attempt count +| Model | Evaluated tasks | Pass@1 | Pass@2 | Pass@4 | Pass@8 | +| --- | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | 5,317 | 10.71% | 15.88% | 21.98% | 28.44% | -You can select tasks by how often they were solved, for example all tasks solved exactly once: +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. The trace export contains all eight attempts per task, including 717 attempts without a raw verifier reward. The plot uses recorded successes out of eight; no missing verifier grades are asserted to be failures. Selection metadata still carries earlier incomplete-status labels for 440 tasks. -```python -import pyarrow.parquet as pq +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)]) -``` +## Files and usage + +`tasks.parquet` contains 1,512 kept tasks. -## Traces +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. -`traces/` holds all 42,536 attempts, successful and failed, from every evaluated task, one row per attempt (Zstandard-compressed Parquet). Columns include `task_id`, `model`, `attempt_id`, `reward`, `verifier_reward`, `retained_v2` and the full `trajectory` and `result` JSON. +Success-count columns: `instruct_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. ```python -import json -import pyarrow.dataset as ds +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -traces = ds.dataset("traces", format="parquet") -failed = traces.to_table(filter=ds.field("reward") == 0) -conversation = json.loads(failed["trajectory"][0].as_py()) +path = hf_hub_download( + "FWeindel/validated-tasks", + "calibforge_v2/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("instruct_successes", "=", 1)]) ``` -Trajectory statistics (plots, command distribution) are in [trace_stats](trace_stats/README.md). +Published task payloads differ from the evaluated payloads by instruction suffix/whitespace changes; 14 kept tasks also predate later instruction repairs. These results do not establish evaluation of the published bytes. + +The optional `tasks_solved_ctx25000_out8192_no_summary.parquet` contains 1,258 existing v2 tasks with at least one successful trace whose recorded steps use at most 25,000 prompt-plus-output tokens, fewer than 8,192 output tokens, and explicitly zero summaries. It is an observed-trace subset, not a new evaluation with enforced limits. Recorded steps can aggregate requests; missing usage or unknown summary counts do not qualify. + +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. -Not yet checked with the repository-leak checks added on 2026-10-17 (leftover git history and file timestamps that could reveal the solution). +## Trace statistics -## Runtime +Of 42,536 recorded attempts, 42,536 have saved conversations and 0 do not. 717 attempts have no raw verifier reward. Missing measurements remain missing. -- **Versions:** Apptainer **1.5.4**, Python **3.12.14**, Harbor **0.8.1**. -- **Implementation:** [Marin Harbor `7faf878c14b6`](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) + [our runtime patches](https://github.com/MLI-lab/OT-next-data/tree/main/harbor_patches). +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -| Model | Mean time per attempt | Median time per attempt | Attempts measured | -| --- | ---: | ---: | ---: | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 470.25 s | 186.52 s | 42,536 | +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -`patch.py` reproduces this release from the pinned v1 Parquet and the recorded counts. +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/calibforge_v2/plots/readme-pass-at-k.png b/calibforge_v2/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..bcfa081f272de5c215fa86d05813b084562e9938 --- /dev/null +++ b/calibforge_v2/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0baae781ba662dae4384694081ed011b99c4e35b33cd18581795c08ff213c0c +size 43119 diff --git a/calibforge_v2/trace_stats/README.md b/calibforge_v2/trace_stats/README.md index b4f4d287ea3cf827497371cc49dd429543233592..10410699a236f87be4079f6d037df7fa3e71caf8 100644 --- a/calibforge_v2/trace_stats/README.md +++ b/calibforge_v2/trace_stats/README.md @@ -1,60 +1,55 @@ -# Trajectory statistics +# CalibForge v2 trace statistics -42,536 attempts across 5,317 tasks of CalibForge v2, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3-30B-A3B-Instruct-2507 | -| --- | ---: | -| Recorded attempts | 42,536 | -| Median agent turns | 13 | -| Median input tokens per attempt | 109,350 | -| Median output tokens per attempt | 7,699 | -| Median agent runtime (seconds) | 159 | -| Median verifier runtime (seconds) | 9 | +Of 42,536 recorded attempts, 42,536 have saved conversations and 0 do not. 717 attempts have no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -![Trace measurements](plots/all-models-trace-distributions.png) +Token totals sum calls across an attempt and are not the context window. -![Trace outcomes](plots/all-models-trace-outcomes.png) +Measurements cover 5,317 tasks, including kept and unselected tasks, with each attempt counted once. -![Most frequent commands](plots/all-models-trace-commands.png) +## Evaluation settings -Commands are counted as submitted, not as successful executions. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 8 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | -Per-model plots: [Qwen/Qwen3-30B-A3B-Instruct-2507 commands](plots/qwen3-30b-instruct-2507-trace-commands.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 distributions](plots/qwen3-30b-instruct-2507-trace-distributions.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 outcomes](plots/qwen3-30b-instruct-2507-trace-outcomes.png). +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Data +**Time limits:** agent 900, 1,200 seconds; verifier 900 seconds (sampled configurations; limits vary by task where multiple values are shown). -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +**Output-limit evidence:** The 2 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -Attempts without a saved conversation: 0; without a verifier reward: 717. Their missing measurements stay empty. +## Measurements -## Evaluation settings +### Qwen/Qwen3-30B-A3B-Instruct-2507 -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 42,536 | 13 | 76 | +| Input tokens per attempt | 42,536 | 109,350 | 1,244,293 | +| Output tokens per attempt | 42,536 | 7,699 | 76,640 | +| Largest recorded step, prompt plus output tokens | 42,508 | 17,125 | 37,095 | +| Agent runtime (seconds) | 42,536 | 159.23 | 1,200 | +| Verifier runtime (seconds) | 42,277 | 8.55 | 15.78 | +| Whole-attempt runtime (seconds) | 42,536 | 186.51 | 1,219.42 | -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. - -The table reports settings found in **2 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](../metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +## Plots -## Token usage and summarization +![Trace measurements](plots/all-models-trace-distributions.png) -All published traces are included, covering kept and excluded tasks and every verifier outcome. One trace means one task attempt. +![Trace outcomes](plots/all-models-trace-outcomes.png) -![Output, summarization, and context distributions](plots/all-models-tokens-by-outcome.png) +![Most frequent commands](plots/all-models-trace-commands.png) -- **Output per recorded agent step:** successful, graded unsuccessful, and ungraded attempts shown separately. Each step contributes once. -- **Summarizations per trace:** one combined distribution across all traces with a recorded count, including zero. -- **Peak context per trace:** the largest recorded step prompt + output count in each trace, separated into traces with and without summarization. Unknown summary counts are kept separate when present. Each trace contributes once, regardless of its length or outcome. +![Output, summarization and recorded context](plots/all-models-tokens-by-outcome.png) -[Detailed statistics](data/tokens-by-outcome.json) retain sample counts and missing measurements. Missing summary counts are not zero. ATIF steps can aggregate multiple requests, so recorded context does not establish an exact per-call context limit. +Output per recorded agent step is separated by verifier outcome. Summarizations are counted per trace, including zero when recorded. Peak recorded context is shown per trace with and without summarization; unknown counts stay separate. diff --git a/calibforge_v3/README.md b/calibforge_v3/README.md index 63db7477ba111162a542e21ba32bcebd95be372d..daf0e1e594f3b8a4addffbc371f38b527b30b7af 100644 --- a/calibforge_v3/README.md +++ b/calibforge_v3/README.md @@ -2,99 +2,92 @@ ## Dataset stats -- **Source v1 tasks:** 5,317 -- **Kept in v3:** 2,287 newly solved tasks -- **Available separately in v2:** 1,512 -- **Combined v2 + v3:** 3,799 unique solved tasks -- **Still unsolved:** 1,518 -- **New evaluation:** 3,805 previously unsolved tasks, 2 attempts each, 7,610 attempts -- **Saved trajectories:** 7,610, successful and failed -- **Changed task payloads:** 0 relative to the evaluated source +- **Tasks:** 3,805 had no successful attempt in v2; all 3,805 evaluated for v3; 2,287 newly kept; 1,518 still without success. +- **Teacher evaluation:** Qwen/Qwen3.8-27B (thinking on), 2 attempts per task (pass@2). +- **Coverage:** 3,805 tasks evaluated; 7,610 evaluation attempts. +- **Saved conversations:** 7,610 of 7,610 published attempt records, including successful and unsuccessful attempts. -CalibForge contains terminal tasks across software engineering, debugging, data processing, scientific computing, system administration and related domains. V3 contains only tasks newly solved by `Qwen/Qwen3.8-27B` after the earlier evaluation. Tasks already solved in [v2](../calibforge_v2/README.md) are not duplicated. This is a harder subset relative to that earlier evaluation, not a claim of intrinsic difficulty for every model. - -## Evaluation - -`Qwen/Qwen3.8-27B` (thinking enabled) used Terminus-2 on the 3,805 tasks left unsolved by v2. The four-task pilot and 3,801-task full run were disjoint. All tasks have exactly two terminal attempts. The task source is `calibforge_v1` at revision `21e998257e3fb6e5396596e921d507d69684d6bf`. Published task payload hashes match the evaluated selection. Historical results are in the v2 README. +We keep tasks with at least one verifier-successful attempt from the new model. `tasks.parquet` contains only newly kept tasks; tasks already kept in v2 remain separate. ## Evaluation settings -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. - -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3.8-27B | 32768 | 8192 | Top-level only in all 7,610 trial configs; explicit cap not forwarded, actual fallback unknown | on | {"min_p": 0.0, "presence_penalty": 0.0, "repetition_penalty": 1.0, "temperature": 1.0, "top_k": 20, "top_p": 0.95} | +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | On | 2 | 1.0 | 0.95 | 20 | 32,768 | 8,192 | min_p=0; presence_penalty=0; repetition_penalty=1 | -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -All 7,610 saved trial configurations were inspected. The intended 8,192-token response limit was stored as top-level `max_tokens`; the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. The actual fallback output cap is unknown. The configured 32,768-token context is not proof of per-request enforcement. Recorded step metrics can aggregate calls or retries; peak recorded step usage is not an independent server-limit measurement. +**Time limits:** agent 120 to 12,000 seconds; verifier 60 to 12,000 seconds (all saved configurations; limits vary by task where multiple values are shown). -Agent time limits range from 120 to 12,000 seconds and verifier limits from 60 to 12,000 seconds, according to each task. [Settings and timeout groups](metadata/evaluation-settings-audit.json) contain the exhaustive audit. +**Output-limit evidence:** All 7,610 saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. -## Results on the previously unsolved subset +## Results -| Model | Evaluated tasks | Attempts | Tasks solved | Pass@1 | Pass@2 | -| --- | ---: | ---: | ---: | ---: | ---: | -| Qwen/Qwen3.8-27B (thinking) | 3,805 | 7,610 | 2,287 | 48.69% | 60.11% | +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. Only the new v3 evaluation is included. -| Successful attempts | Tasks | +| Successful attempts | Qwen/Qwen3.8-27B (3,805 tasks) | | --- | ---: | | 0/2 | 1,518 | | 1/2 | 869 | | 2/2 | 1,418 | -![Successful attempts per task](trace_stats/plots/success-counts.png) +| Model | Evaluated tasks | Pass@1 | Pass@2 | +| --- | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | 3,805 | 48.69% | 60.11% | -These rates use all 3,805 evaluated tasks, not the 2,287 success-selected tasks published in `tasks.parquet`. There were 3,705 successful attempts. The recorded `terminal-errors-zero-v2` policy counts 186 terminal errors as zero while preserving their null raw verifier rewards; 7,424 attempts have verifier rewards. Errors are not passed validations. +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. -## Traces and statistics +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -`traces/` contains all 7,610 new attempts, including failures and attempts on still-unsolved tasks, as compressed Parquet. Each row includes `task_id`, `model`, `attempt_id`, `content_sha256`, `reward`, `verifier_reward`, `reward_origin`, `retained_v3`, and full `trajectory` and `result` JSON. `retained_v3` identifies the incremental task set. Historical traces remain in v2. +## Files and usage -Every new attempt has a saved trajectory. No quality filter was applied. A saved conversation is not necessarily suitable for training, and an agent timeout can still be followed by a successful verifier result. [Recorded errors and release checks](metadata/release-verification.json) preserve counts. +`tasks.parquet` contains 2,287 newly kept tasks only. Tasks already kept in v2 are not duplicated. -| New evaluation measurement | Median | -| --- | ---: | -| Agent turns | 11 | -| Input tokens per attempt | 107,048 | -| Output tokens per attempt | 33,272 | -| Agent runtime | 725.31 s | -| Verifier runtime | 11.59 s | -| Whole-attempt runtime | 751.61 s | +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. -Token totals sum calls across each attempt and are not context-window measurements. Mean whole-attempt runtime was 874.11 seconds. [Trace statistics](trace_stats/README.md) include plots, outcome and command distributions, and per-task and per-attempt measurements. +Success-count columns: `qwen38_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. -## Use v2 and v3 together +`archive.parquet` preserves excluded task payloads and reasons, including earlier exclusions where applicable. ```python from huggingface_hub import hf_hub_download -import pyarrow as pa import pyarrow.parquet as pq +path = hf_hub_download( + "FWeindel/validated-tasks", + "calibforge_v3/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("qwen38_successes", "=", 1)]) +``` + +Combine the 1,512 tasks in `calibforge_v2/tasks.parquet` with these 2,287 newly kept tasks for 3,799 unique tasks. Allow for different columns and retain each source file's image manifest when staging images. + +```python +import pyarrow as pa + def load(version): - path = hf_hub_download("FWeindel/validated-tasks", - f"calibforge_{version}/tasks.parquet", repo_type="dataset") + path = hf_hub_download( + "FWeindel/validated-tasks", + f"calibforge_{version}/tasks.parquet", + repo_type="dataset", + ) return pq.read_table(path) v2, v3 = load("v2"), load("v3") assert set(v2["path"].to_pylist()).isdisjoint(v3["path"].to_pylist()) combined = pa.concat_tables([v2, v3], promote_options="permissive") -assert combined.num_rows == 3799 ``` -The example combines task rows; each source file retains its own image manifest. Use the source manifests when staging images. V3 adds `qwen38_successes`, `qwen38_attempts`, and `kept_from_v2` (false for every v3 row). To select tasks solved exactly once, filter `qwen38_successes == 1`. +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. -## Files and provenance +## Trace statistics -- [tasks.parquet](tasks.parquet): 2,287 newly solved tasks. -- [archive.parquet](archive.parquet): 1,518 evaluated tasks without a success; these are not training task rows. -- [Selection](metadata/selection.json): membership, evaluated payload hashes and success counts for all 3,805 tasks. -- [Trace manifest](traces/manifest.json): attempt counts and file checksums. -- [Pilot contract](metadata/pilot-contract.json) and [full contract](metadata/full-contract.json): frozen evaluation configuration. -- [patch.py](patch.py): reproduce the incremental task selection from the pinned source. +Of 7,610 recorded attempts, 7,610 have saved conversations and 0 do not. 186 attempts have no raw verifier reward. Missing measurements remain missing. -## Runtime and limitations +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -The run used one H200 and 32 allocated CPUs per shard, with four requested concurrent trials. Saved reports identify Apptainer 1.5.4, Python 3.12.14, Harbor 0.8.1, and Marin Harbor revision `7faf878c14b6d72737579ec936ebf5f74ba3194d` with the captured runtime patches. +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. Historical attempts remain in v2. -This release adds teacher results; it does not establish completion of all current datasource-validation checks. The earlier repository-history and timestamp leak-check limitation remains unresolved. See v2 for inherited validation limitations and the settings section for the historical output-cap issue. +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/calibforge_v3/plots/readme-pass-at-k.png b/calibforge_v3/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..c505e7409a0ee87fa19b4e988cbecf2f2a8c0376 --- /dev/null +++ b/calibforge_v3/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afeeedd72fd3844e0b9164c219fab814e75600a99d9d090b06e7304d3e01d0f8 +size 38676 diff --git a/calibforge_v3/trace_stats/README.md b/calibforge_v3/trace_stats/README.md index 52a04e8328e4f5535e2903f399237c6ea7fb4124..bc2ed2a3555fd404a95a7ca6f3b6cbec1171a1d5 100644 --- a/calibforge_v3/trace_stats/README.md +++ b/calibforge_v3/trace_stats/README.md @@ -1,65 +1,55 @@ -# Trajectory statistics +# CalibForge v3 trace statistics -7,610 attempts across 3,805 tasks of calibforge_v3, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3.8-27B | -| --- | ---: | -| Recorded attempts | 7,610 | -| Median agent turns | 11 | -| Median input tokens per attempt | 107,048 | -| Median output tokens per attempt | 33,272 | -| Median agent runtime (seconds) | 725 | -| Median verifier runtime (seconds) | 12 | +Of 7,610 recorded attempts, 7,610 have saved conversations and 0 do not. 186 attempts have no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots - -![Trace measurements](plots/all-models-trace-distributions.png) - -![Trace outcomes](plots/all-models-trace-outcomes.png) - -![Most frequent commands](plots/all-models-trace-commands.png) +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. Historical attempts remain in v2. -Commands are counted as submitted, not as successful executions. +Token totals sum calls across an attempt and are not the context window. -Per-model plots: [Qwen/Qwen3.8-27B commands](plots/qwen3.8-27b-trace-commands.png), [Qwen/Qwen3.8-27B distributions](plots/qwen3.8-27b-trace-distributions.png), [Qwen/Qwen3.8-27B outcomes](plots/qwen3.8-27b-trace-outcomes.png). +Measurements cover 3,805 tasks, including kept and unselected tasks, with each attempt counted once. -## Data - -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +## Evaluation settings -Attempts without a saved conversation: 0; without a verifier reward: 186. Their missing measurements stay empty. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | On | 2 | 1.0 | 0.95 | 20 | 32,768 | 8,192 | min_p=0; presence_penalty=0; repetition_penalty=1 | +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -Evaluation budgets: see the [dataset evaluation setup](../README.md). Per-call context and output-limit enforcement are not established by token-usage statistics. +**Time limits:** agent 120 to 12,000 seconds; verifier 60 to 12,000 seconds (all saved configurations; limits vary by task where multiple values are shown). +**Output-limit evidence:** All 7,610 saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. -## Token usage and summarization +## Measurements -All published traces are included, covering kept and excluded tasks and every verifier outcome. One trace means one task attempt. +### Qwen/Qwen3.8-27B -![Output, summarization, and context distributions](plots/all-models-tokens-by-outcome.png) +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 7,610 | 11 | 21 | +| Input tokens per attempt | 7,610 | 107,048 | 299,871 | +| Output tokens per attempt | 7,610 | 33,272 | 80,274 | +| Largest recorded step, prompt plus output tokens | 7,594 | 24,531 | 53,797 | +| Agent runtime (seconds) | 7,610 | 725.31 | 1,800 | +| Verifier runtime (seconds) | 7,446 | 11.59 | 20 | +| Whole-attempt runtime (seconds) | 7,610 | 751.61 | 1,824.23 | -- **Output per recorded agent step:** successful, graded unsuccessful, and ungraded attempts shown separately. Each step contributes once. -- **Summarizations per trace:** one combined distribution across all traces with a recorded count, including zero. -- **Peak context per trace:** the largest recorded step prompt + output count in each trace, separated into traces with and without summarization. Unknown summary counts are kept separate when present. Each trace contributes once, regardless of its length or outcome. +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -[Detailed statistics](data/tokens-by-outcome.json) retain sample counts and missing measurements. Missing summary counts are not zero. ATIF steps can aggregate multiple requests, so recorded context does not establish an exact per-call context limit. -## Evaluation settings +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +## Plots -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3.8-27B | 32768 | 8192 | Top-level only in all 7,610 trial configs; explicit cap not forwarded, actual fallback unknown | on | {"min_p": 0.0, "presence_penalty": 0.0, "repetition_penalty": 1.0, "temperature": 1.0, "top_k": 20, "top_p": 0.95} | +![Trace measurements](plots/all-models-trace-distributions.png) -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +![Trace outcomes](plots/all-models-trace-outcomes.png) -All 7,610 saved trial configurations were inspected. The intended 8,192-token response limit was stored as top-level `max_tokens`; the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. The actual fallback output cap is unknown. The configured 32,768-token context is not proof of per-request enforcement. Recorded step metrics can aggregate calls or retries; peak recorded step usage is not an independent server-limit measurement. +![Most frequent commands](plots/all-models-trace-commands.png) -Agent time limits range from 120 to 12,000 seconds and verifier limits from 60 to 12,000 seconds, according to each task. [Settings and timeout groups](../metadata/evaluation-settings-audit.json) contain the exhaustive audit. +![Output, summarization and recorded context](plots/all-models-tokens-by-outcome.png) -Earlier evaluation statistics are in [v2](../../calibforge_v2/trace_stats/README.md). +Output per recorded agent step is separated by verifier outcome. Summarizations are counted per trace, including zero when recorded. Peak recorded context is shown per trace with and without summarization; unknown counts stay separate. diff --git a/crosscodeeval_v2/README.md b/crosscodeeval_v2/README.md index b15661840d878e27f9395560100ee276bc6cd734..2a15b83aaa4a64f265ca2e1f925118ed0b9966c6 100644 --- a/crosscodeeval_v2/README.md +++ b/crosscodeeval_v2/README.md @@ -2,39 +2,33 @@ ## Dataset stats -- **v1 tasks evaluated:** 6,710 -- **Kept:** 863 (solved at least once by either model) -- **Excluded:** 5,847 (no recorded success) -- **Attempts:** pass@16 with Qwen/Qwen3-30B-A3B-Instruct-2507 and Qwen/Qwen3-30B-A3B (thinking) (214,720 in total) +- **Tasks:** 6,710 in v1; 863 kept in v2. +- **Teacher evaluation:** Qwen/Qwen3-30B-A3B-Instruct-2507 (thinking off), 16 attempts per task (pass@16); Qwen/Qwen3-30B-A3B (thinking on), 16 attempts per task (pass@16). +- **Coverage:** 6,710/6,710 tasks evaluated; 214,720 evaluation attempts. +- **Saved conversations:** 214,719 of 214,720 published attempt records, including successful and unsuccessful attempts. -## Evaluation settings - -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +We keep tasks with at least one verifier-successful attempt from any evaluated model. -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | -| Qwen/Qwen3-30B-A3B | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | on | {"min_p": 0, "temperature": 0.6, "top_k": 20, "top_p": 0.95} | +## Evaluation settings -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 16 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | +| Qwen/Qwen3-30B-A3B | Terminus-2 | On | 16 | 0.6 | 0.95 | 20 | 32,768 | 8,192 | min_p=0 | -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -The table reports settings found in **6 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +**Time limits:** agent 900 seconds; verifier 720 seconds (sampled configurations; limits vary by task where multiple values are shown). -## Pass@16 filtering +**Output-limit evidence:** The 6 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -| Model | Tasks evaluated | Attempts | Tasks solved at least once | -| --- | ---: | ---: | ---: | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 6,710 | 107,360 | 641 | -| Qwen/Qwen3-30B-A3B (thinking) | 6,710 | 107,360 | 627 | -| Either model | 6,710 | 214,720 | 863 | +## Results -### Successful attempts among kept tasks +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. -| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 tasks | Qwen/Qwen3-30B-A3B (thinking) tasks | +| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 (6,710 tasks) | Qwen/Qwen3-30B-A3B (6,710 tasks) | | --- | ---: | ---: | -| 0/16 | 222 | 236 | +| 0/16 | 6,069 | 6,083 | | 1/16 | 222 | 193 | | 2/16 | 70 | 94 | | 3/16 | 57 | 64 | @@ -52,36 +46,46 @@ The table reports settings found in **6 sampled saved trial configurations**, no | 15/16 | 7 | 7 | | 16/16 | 13 | 15 | -## Select tasks by successful-attempt count +| Model | Evaluated tasks | Pass@1 | Pass@2 | Pass@4 | Pass@8 | Pass@16 | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | 6,710 | 2.80% | 4.19% | 5.79% | 7.54% | 9.55% | +| Qwen/Qwen3-30B-A3B | 6,710 | 2.69% | 4.04% | 5.65% | 7.45% | 9.34% | -You can select tasks by how often they were solved, for example all tasks Qwen/Qwen3-30B-A3B-Instruct-2507 solved exactly once: +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. -```python -import pyarrow.parquet as pq +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)]) -``` +## Files and usage + +`tasks.parquet` contains 863 kept tasks. -## Traces +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. -`traces/` holds all 214,720 recorded attempts, successful and failed, from every evaluated task, one row per attempt (Zstandard-compressed Parquet). Columns include `task_id`, `model`, `attempt_id`, `reward`, `verifier_reward`, `retained_v2` and the full `trajectory` and `result` JSON. 1 attempts have no saved conversation (`trajectory_found=false`). +Success-count columns: `instruct_successes`, `thinking_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. ```python -import json -import pyarrow.dataset as ds +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -traces = ds.dataset("traces", format="parquet") -failed = traces.to_table(filter=ds.field("reward") == 0) -conversation = json.loads(failed["trajectory"][0].as_py()) +path = hf_hub_download( + "FWeindel/validated-tasks", + "crosscodeeval_v2/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("instruct_successes", "=", 1)]) ``` -Trajectory statistics (plots, command distribution) are in [trace_stats](trace_stats/README.md). +The evaluated payloads omit the timeout/anti-cheating instruction suffix preserved in the published task files. + +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. + +## Trace statistics -Not yet checked with the repository-leak checks added on 2026-10-17 (leftover git history and file timestamps that could reveal the solution). +Of 214,720 recorded attempts, 214,719 have saved conversations and 1 do not. 2 attempts have no raw verifier reward. Missing measurements remain missing. -## Runtime +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -- **Versions:** Apptainer **1.5.4**, Python **3.12.14**, Harbor **0.8.1**. -- **Implementation:** [Marin Harbor `7faf878c14b6`](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) + [our runtime patches](https://github.com/MLI-lab/OT-next-data/tree/main/harbor_patches). +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -`patch.py` rebuilds `tasks.parquet` from the current v1 tasks and the recorded results, both pinned to revision `a43f5635`. +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/crosscodeeval_v2/plots/readme-pass-at-k.png b/crosscodeeval_v2/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..a8150bc0ddba1c1a832040607fed2f36094e1d78 --- /dev/null +++ b/crosscodeeval_v2/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f6508c55b9c0ac0ba57782cdd5c5601bb032620a32e257dbb715f87d4524c48 +size 50961 diff --git a/crosscodeeval_v2/trace_stats/README.md b/crosscodeeval_v2/trace_stats/README.md index 8f7352e55967a54e97f9bc4f97883ee670c1d7a2..af2ee6b7ca59f0af9401db48986b08c4f05f005d 100644 --- a/crosscodeeval_v2/trace_stats/README.md +++ b/crosscodeeval_v2/trace_stats/README.md @@ -1,49 +1,64 @@ -# Trajectory statistics +# CrossCodeEval v2 trace statistics -214,720 attempts across 6,710 tasks of crosscodeeval_v2, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3-30B-A3B-Instruct-2507 | Qwen/Qwen3-30B-A3B (thinking) | -| --- | ---: | ---: | -| Recorded attempts | 107,360 | 107,360 | -| Median agent turns | 2 | 2 | -| Median input tokens per attempt | 3,865 | 3,723 | -| Median output tokens per attempt | 393 | 1,504 | -| Median agent runtime (seconds) | 6 | 20 | -| Median verifier runtime (seconds) | 1 | 1 | +Of 214,720 recorded attempts, 214,719 have saved conversations and 1 do not. 2 attempts have no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -![Trace measurements](plots/all-models-trace-distributions.png) +Token totals sum calls across an attempt and are not the context window. -![Trace outcomes](plots/all-models-trace-outcomes.png) +Measurements cover 6,710 tasks, including kept and unselected tasks, with each attempt counted once. -![Most frequent commands](plots/all-models-trace-commands.png) +## Evaluation settings -Commands are counted as submitted, not as successful executions. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 16 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | +| Qwen/Qwen3-30B-A3B | Terminus-2 | On | 16 | 0.6 | 0.95 | 20 | 32,768 | 8,192 | min_p=0 | -Per-model plots: [Qwen/Qwen3-30B-A3B-Instruct-2507 commands](plots/qwen3-30b-instruct-2507-trace-commands.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 distributions](plots/qwen3-30b-instruct-2507-trace-distributions.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 outcomes](plots/qwen3-30b-instruct-2507-trace-outcomes.png), [Qwen/Qwen3-30B-A3B (thinking) commands](plots/qwen3-30b-thinking-trace-commands.png), [Qwen/Qwen3-30B-A3B (thinking) distributions](plots/qwen3-30b-thinking-trace-distributions.png), [Qwen/Qwen3-30B-A3B (thinking) outcomes](plots/qwen3-30b-thinking-trace-outcomes.png). +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Data +**Time limits:** agent 900 seconds; verifier 720 seconds (sampled configurations; limits vary by task where multiple values are shown). -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +**Output-limit evidence:** The 6 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -Attempts without a saved conversation: 1; without a verifier reward: 2. Their missing measurements stay empty. +## Measurements -## Evaluation settings +### Qwen/Qwen3-30B-A3B-Instruct-2507 + +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 107,359 | 2 | 2 | +| Input tokens per attempt | 107,359 | 3,865 | 7,488 | +| Output tokens per attempt | 107,359 | 393 | 616 | +| Largest recorded step, prompt plus output tokens | 107,338 | 2,247 | 3,575 | +| Agent runtime (seconds) | 107,360 | 6.29 | 9.72 | +| Verifier runtime (seconds) | 107,360 | 1.24 | 1.38 | +| Whole-attempt runtime (seconds) | 107,360 | 13.72 | 34.03 | -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +### Qwen/Qwen3-30B-A3B -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | -| Qwen/Qwen3-30B-A3B | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | on | {"min_p": 0, "temperature": 0.6, "top_k": 20, "top_p": 0.95} | +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 107,360 | 2 | 2 | +| Input tokens per attempt | 107,360 | 3,723 | 7,102 | +| Output tokens per attempt | 107,360 | 1,504 | 2,906 | +| Largest recorded step, prompt plus output tokens | 107,330 | 2,917 | 4,658 | +| Agent runtime (seconds) | 107,360 | 19.77 | 38.25 | +| Verifier runtime (seconds) | 107,358 | 1.23 | 1.38 | +| Whole-attempt runtime (seconds) | 107,360 | 29.41 | 70 | -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -The table reports settings found in **6 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](../metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +## Plots + +![Trace measurements](plots/all-models-trace-distributions.png) + +![Trace outcomes](plots/all-models-trace-outcomes.png) + +![Most frequent commands](plots/all-models-trace-commands.png) diff --git a/endless_v2/README.md b/endless_v2/README.md index f3f0ba5eea4fd20fa0cb7bebdb7e9ae929012190..bd9e049ec3119656a11cc97e8a5da6e002e2cdb5 100644 --- a/endless_v2/README.md +++ b/endless_v2/README.md @@ -2,86 +2,94 @@ ## Dataset stats -- **v1 tasks:** 2,144 (from 2,492 original tasks) -- **Tasks evaluated:** 2,141 with both models -- **Kept:** 1,958 (solved at least once by either model), including 635 added by Qwen/Qwen3-Coder-30B-A3B-Instruct -- **Excluded:** 186 (183 with no recorded success; 3 with no recorded attempt) -- **Attempts:** 25,692 in total, 8,564 Qwen/Qwen3-30B-A3B-Instruct-2507 and 17,128 Qwen/Qwen3-Coder-30B-A3B-Instruct +- **Tasks:** 2,144 in v1; 1,958 kept in v2. +- **Teacher evaluation:** Qwen/Qwen3-30B-A3B-Instruct-2507 (thinking off), 4 attempts per task (pass@4); Qwen/Qwen3-Coder-30B-A3B-Instruct (thinking off), 8 attempts per task (pass@8). +- **Coverage:** 2,141/2,144 tasks evaluated; 25,692 evaluation attempts. +- **Saved conversations:** 25,691 of 25,692 published attempt records, including successful and unsuccessful attempts. + +We keep tasks with at least one verifier-successful attempt from any evaluated model. ## Evaluation settings -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 4 | 0.7 | 0.8 | 20 | 65,536 | 16,384 | min_p=0 | +| Qwen/Qwen3-Coder-30B-A3B-Instruct | Terminus-2 | Off | 8 | 0.7 | 0.8 | 20 | 65,536 | 16,384 | repetition_penalty=1.05 | -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 65536 | 16384 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | -| Qwen/Qwen3-Coder-30B-A3B-Instruct | 65536 | 16384 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"repetition_penalty": 1.05, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +**Time limits:** agent 300 seconds; verifier 300 seconds (sampled configurations; limits vary by task where multiple values are shown). -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +**Output-limit evidence:** The 2 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -The table reports settings found in **2 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +## Results -## Pass@k filtering +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. -| Model | Tasks evaluated | Attempts per task | Tasks solved at least once | Pass@1 | Pass@4 | Pass@8 | -| --- | ---: | ---: | ---: | ---: | ---: | ---: | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 2,141 | 4 | 1,323 | 42.73% | 61.79% | n/a | -| Qwen/Qwen3-Coder-30B-A3B-Instruct | 2,141 | 8 | 1,924 | 66.39% | 85.43% | 89.86% | +Both models evaluated the same 2,141 tasks; 3 v1 tasks had no recorded attempt and are excluded from these tables. The eight-attempt evaluation combines two four-attempt collections. -Both models ran on the same 2,141 tasks, including tasks already solved by Qwen/Qwen3-30B-A3B-Instruct-2507. Qwen/Qwen3-Coder-30B-A3B-Instruct's eight attempts came from two four-attempt collections. All measured tasks have their full recorded attempt count; three v1 tasks have no recorded attempt. Rates cover all measured tasks before filtering. Pass@k uses the unbiased estimate from each model's full attempt budget; Qwen/Qwen3-30B-A3B-Instruct-2507 has no pass@8 measurement. +| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 (2,141 tasks) | +| --- | ---: | +| 0/4 | 818 | +| 1/4 | 288 | +| 2/4 | 254 | +| 3/4 | 261 | +| 4/4 | 520 | -Keep a task if either model solved it at least once. Of the 1,958 kept tasks, 1,289 were solved by both models, 635 only by Qwen/Qwen3-Coder-30B-A3B-Instruct and 34 only by Qwen/Qwen3-30B-A3B-Instruct-2507. The 348 original tasks excluded before v1 remain excluded. +| Successful attempts | Qwen/Qwen3-Coder-30B-A3B-Instruct (2,141 tasks) | +| --- | ---: | +| 0/8 | 217 | +| 1/8 | 121 | +| 2/8 | 109 | +| 3/8 | 127 | +| 4/8 | 138 | +| 5/8 | 171 | +| 6/8 | 239 | +| 7/8 | 342 | +| 8/8 | 677 | -59 kept tasks were evaluated on an earlier version of the task; `tasks.parquet` has the current v1 version. Trace and pass-count hashes identify the evaluated payload, and selection metadata also records each current payload hash. +| Model | Evaluated tasks | Pass@1 | Pass@2 | Pass@4 | Pass@8 | +| --- | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | 2,141 | 42.73% | 53.09% | 61.79% | Not evaluated | +| Qwen/Qwen3-Coder-30B-A3B-Instruct | 2,141 | 66.39% | 78.14% | 85.43% | 89.86% | -### Successful attempts among kept tasks +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. -Counts below cover all 1,958 retained tasks. A zero for one model means the other model solved that task. +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 tasks (out of 4 attempts) | Qwen/Qwen3-Coder-30B-A3B-Instruct tasks (out of 8 attempts) | -| --- | ---: | ---: | -| 0 | 635 | 34 | -| 1 | 288 | 121 | -| 2 | 254 | 109 | -| 3 | 261 | 127 | -| 4 | 520 | 138 | -| 5 | n/a | 171 | -| 6 | n/a | 239 | -| 7 | n/a | 342 | -| 8 | n/a | 677 | +## Files and usage -## Select tasks by successful-attempt count +`tasks.parquet` contains 1,958 kept tasks. -```python -import pyarrow.parquet as pq - -# Tasks Qwen/Qwen3-30B-A3B-Instruct-2507 solved exactly once in four attempts. -instruct_subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)]) -# Tasks Qwen/Qwen3-Coder-30B-A3B-Instruct solved exactly once in eight attempts. -coder_subset = pq.read_table("tasks.parquet", filters=[("coder_successes", "=", 1)]) -``` +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. -Use `instruct_target_attempts` (4) and `coder_target_attempts` (8) for the model budgets. The legacy `target_attempts_per_model` column remains 4 for compatibility. [Selection metadata](metadata/selection.json) covers all 2,144 tasks; [pass counts](pass_counts.parquet) and [pass@k results](metadata/pass_at_k.json) preserve the per-model results. Excluded v1 task payloads are in [archive.parquet](archive.parquet). +Success-count columns: `instruct_successes`, `coder_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. -## Traces - -[traces/attempts.parquet](traces/attempts.parquet) contains all **25,692 recorded attempts** from both models, successful and failed, across kept and excluded tasks. Each row includes the model, task ID, attempt ID, evaluated content hash, reward, raw verifier reward, retention flag, result JSON and trajectory JSON when available. All 8,564 Qwen/Qwen3-30B-A3B-Instruct-2507 conversations and 17,127 of 17,128 Qwen/Qwen3-Coder-30B-A3B-Instruct conversations are available; the remaining Qwen/Qwen3-Coder-30B-A3B-Instruct attempt retains its result record. +`archive.parquet` preserves excluded task payloads and reasons, including earlier exclusions where applicable. ```python -import pyarrow.dataset as ds +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -traces = ds.dataset("traces/attempts.parquet", format="parquet") -coder = traces.to_table(filter=ds.field("model") == "Qwen/Qwen3-Coder-30B-A3B-Instruct") +path = hf_hub_download( + "FWeindel/validated-tasks", + "endless_v2/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("instruct_successes", "=", 1)]) ``` -[Trace statistics](trace_stats/README.md) show separate Qwen/Qwen3-30B-A3B-Instruct-2507 and Qwen/Qwen3-Coder-30B-A3B-Instruct summaries and plots, plus the combined overview. They include every recorded attempt, not just successful or retained tasks. Missing conversations and verifier grades remain recorded as missing. +59 kept tasks have evaluation evidence from an earlier task version; the published file contains the updated v1 payloads. + +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. + +## Trace statistics -Not yet checked with the repository-leak checks added on 2026-10-17 (leftover git history and file timestamps that could reveal the solution). +Of 25,692 recorded attempts, 25,691 have saved conversations and 1 do not. 15 attempts have no raw verifier reward. Missing measurements remain missing. -## Runtime and reproduction +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -Teacher runs used one H200, 32 CPUs per job and concurrency 10. Each collection used fifteen jobs taking approximately 55 to 79 minutes each. Environment and image versions are inherited from [v1](../endless_v1/README.md). +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -`patch.py` rebuilds the task, archive and pass-count Parquets from pinned current v1 payloads and the recorded selection. This release incorporates the Qwen/Qwen3-30B-A3B-Instruct-2507 and Qwen/Qwen3-Coder-30B-A3B-Instruct evidence from [PR #22](https://huggingface.co/datasets/FWeindel/validated-tasks/discussions/22) while preserving subsequent v1 repairs. +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/endless_v2/plots/readme-pass-at-k.png b/endless_v2/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..aa8595efe026979c4a354480c24b9578ab93f54a --- /dev/null +++ b/endless_v2/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f55dc9c923edfb699c0aec31b207fa56c0e29f09f9e767fc7a851c8353ba494 +size 57915 diff --git a/endless_v2/trace_stats/README.md b/endless_v2/trace_stats/README.md index d7bfc1b9274cd86a0610e45593147013968ca80b..05ab9eabe94f7726c50086eef40db383051ed0ce 100644 --- a/endless_v2/trace_stats/README.md +++ b/endless_v2/trace_stats/README.md @@ -1,64 +1,64 @@ -# Trajectory statistics +# EndlessTerminals v2 trace statistics -25,692 attempts across 2,141 tasks of EndlessTerminals v2, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3-30B-A3B-Instruct-2507 | Qwen/Qwen3-Coder-30B-A3B-Instruct | -| --- | ---: | ---: | -| Recorded attempts | 8,564 | 17,128 | -| Median agent turns | 3 | 19 | -| Median input tokens per attempt | 6,254 | 64,495 | -| Median output tokens per attempt | 1,006 | 2,515 | -| Median agent runtime (seconds) | 14 | 32 | -| Median verifier runtime (seconds) | 3 | 3 | +Of 25,692 recorded attempts, 25,691 have saved conversations and 1 do not. 15 attempts have no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Qwen/Qwen3-30B-A3B-Instruct-2507 plots +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -![Qwen/Qwen3-30B-A3B-Instruct-2507: Trace measurements](plots/qwen3-30b-a3b-instruct-2507-trace-distributions.png) +Token totals sum calls across an attempt and are not the context window. -![Qwen/Qwen3-30B-A3B-Instruct-2507: Trace outcomes](plots/qwen3-30b-a3b-instruct-2507-trace-outcomes.png) +Measurements cover 2,141 tasks, including kept and unselected tasks, with each attempt counted once. -![Qwen/Qwen3-30B-A3B-Instruct-2507: Most frequent commands](plots/qwen3-30b-a3b-instruct-2507-trace-commands.png) - -## Qwen/Qwen3-Coder-30B-A3B-Instruct plots - -![Qwen/Qwen3-Coder-30B-A3B-Instruct: Trace measurements](plots/qwen3-coder-30b-a3b-instruct-trace-distributions.png) +## Evaluation settings -![Qwen/Qwen3-Coder-30B-A3B-Instruct: Trace outcomes](plots/qwen3-coder-30b-a3b-instruct-trace-outcomes.png) +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 4 | 0.7 | 0.8 | 20 | 65,536 | 16,384 | min_p=0 | +| Qwen/Qwen3-Coder-30B-A3B-Instruct | Terminus-2 | Off | 8 | 0.7 | 0.8 | 20 | 65,536 | 16,384 | repetition_penalty=1.05 | -![Qwen/Qwen3-Coder-30B-A3B-Instruct: Most frequent commands](plots/qwen3-coder-30b-a3b-instruct-trace-commands.png) +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -Commands are counted as submitted, not as successful executions. +**Time limits:** agent 300 seconds; verifier 300 seconds (sampled configurations; limits vary by task where multiple values are shown). -Combined plots: [distributions](plots/all-models-trace-distributions.png), [outcomes](plots/all-models-trace-outcomes.png), [commands](plots/all-models-trace-commands.png). +**Output-limit evidence:** The 2 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -## Data +## Measurements -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +### Qwen/Qwen3-30B-A3B-Instruct-2507 -Attempts without a saved conversation: 1; without a verifier reward: 15. Their missing measurements stay empty. +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 8,564 | 3 | 11 | +| Input tokens per attempt | 8,564 | 6,254 | 50,355 | +| Output tokens per attempt | 8,564 | 1,006 | 4,563 | +| Largest recorded step, prompt plus output tokens | 8,404 | 2,774 | 8,356 | +| Agent runtime (seconds) | 8,564 | 13.87 | 169.76 | +| Verifier runtime (seconds) | 8,560 | 3.33 | 3.69 | +| Whole-attempt runtime (seconds) | 8,564 | 27.53 | 182.29 | -| Missing evidence | Qwen/Qwen3-30B-A3B-Instruct-2507 | Qwen/Qwen3-Coder-30B-A3B-Instruct | -| --- | ---: | ---: | -| Saved conversation | 0 | 1 | -| Raw verifier reward | 5 | 10 | +### Qwen/Qwen3-Coder-30B-A3B-Instruct -Statistics describe the recorded evaluation versions. The downloadable task Parquet uses the latest pinned v1 payloads; 59 retained tasks have since changed. +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 17,127 | 19 | 37 | +| Input tokens per attempt | 17,127 | 64,495 | 247,661 | +| Output tokens per attempt | 17,127 | 2,515 | 5,913 | +| Largest recorded step, prompt plus output tokens | 17,127 | 5,699 | 12,137 | +| Agent runtime (seconds) | 17,127 | 32.35 | 77.61 | +| Verifier runtime (seconds) | 17,119 | 3.33 | 3.48 | +| Whole-attempt runtime (seconds) | 17,128 | 45.74 | 91.2 | -## Evaluation settings +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 65536 | 16384 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | -| Qwen/Qwen3-Coder-30B-A3B-Instruct | 65536 | 16384 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"repetition_penalty": 1.05, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +## Plots -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +![Trace measurements](plots/all-models-trace-distributions.png) -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +![Trace outcomes](plots/all-models-trace-outcomes.png) -The table reports settings found in **2 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](../metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +![Most frequent commands](plots/all-models-trace-commands.png) diff --git a/endless_v3/README.md b/endless_v3/README.md index 2f0b213a496d18c4a68b0e5190d2645dcdde98fa..2b0612e9d6f2b178ac559e98d5ab6f49fc573764 100644 --- a/endless_v3/README.md +++ b/endless_v3/README.md @@ -2,107 +2,81 @@ ## Dataset stats -- **v1 tasks:** 2,144 -- **Kept in v3:** 121 newly solved tasks -- **Available separately in v2:** 1,958 -- **Combined v2 + v3:** 2,079 unique tasks -- **Excluded:** 65 (no recorded success from any of the three models) -- **New evaluation:** the 186 tasks excluded from v2 (183 unsolved, 3 never attempted), 2 attempts each, 372 attempts -- **New saved trajectories:** 372, successful and failed +- **Tasks:** 183 had no successful attempt in v2 and 3 were not attempted; all 186 evaluated for v3; 121 newly kept; 65 still without success. +- **Teacher evaluation:** Qwen/Qwen3.8-27B (thinking on), 2 attempts per task (pass@2). +- **Coverage:** 186 tasks evaluated; 372 evaluation attempts. +- **Saved conversations:** 372 of 372 published attempt records, including successful and unsuccessful attempts. -`tasks.parquet` contains only newly solved tasks, preserving their evaluated payloads and recorded result columns. +We keep tasks with at least one verifier-successful attempt from the new model. `tasks.parquet` contains only newly kept tasks; tasks already kept in v2 remain separate. ## Evaluation settings -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | On | 2 | 1.0 | 0.95 | 20 | 32,768 | 8,192 | min_p=0; presence_penalty=0; repetition_penalty=1 | -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3.8-27B | 32768 | 8192 | `llm_call_kwargs.max_tokens` = 8192 in all 372 saved trial configs; 37 pilot request payloads checked | on | {"min_p": 0.0, "presence_penalty": 0.0, "repetition_penalty": 1.0, "temperature": 1.0, "top_k": 20, "top_p": 0.95} | +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -Configured limits are not measurements of token usage. Agent and verifier time limits are 300 s each for every evaluated task. Unlike the v2 runs, this run forwards the output cap to each request; the earlier attempt at this run (job 1065977) did not and was cancelled before it produced results. +**Time limits:** agent 300 seconds; verifier 300 seconds (all 372 saved configurations; limits vary by task where multiple values are shown). -Settings and evidence: [audit](metadata/evaluation-settings-audit.json), [pilot contract](metadata/pilot-contract.json), [full contract](metadata/full-contract.json). +**Output-limit evidence:** All 372 saved trial configurations contain an explicit 8,192-token request cap; the 37 saved pilot request payloads confirm it. Full-run HTTP payloads were not audited. -## Evaluation results +Network isolation was requested, but the run fell back to host networking when network namespace creation failed; agents could have reached the network. -| Model | Tasks evaluated | Attempts | Tasks solved at least once | Pass@1 | Pass@2 | -| --- | ---: | ---: | ---: | ---: | ---: | -| Qwen/Qwen3.8-27B (thinking) | 186 | 372 | 121 | 57.80% | 65.05% | +## Results -These results cover only the new evaluation of the tasks v2 excluded; earlier results are in [v2](../endless_v2/README.md). Four tasks ran in the pilot (job 1085307) and 182 in the full run (job 1090703) with the same model, sampling and limits. +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. Only the new v3 evaluation is included. -### Successful attempts in the new evaluation +The new evaluation includes 183 tasks with no success in v2 and 3 tasks that v2 never attempted. Stage-1 instruction-path recovery was incomplete for 3 pilot tasks. One evaluated task failed the static nproc check and remains unsolved and excluded. -| Successful attempts | Tasks | +| Successful attempts | Qwen/Qwen3.8-27B (186 tasks) | | --- | ---: | | 0/2 | 65 | | 1/2 | 27 | | 2/2 | 94 | -![Qwen/Qwen3.8-27B success distribution](trace_stats/plots/qwen38-success-counts.png) +| Model | Evaluated tasks | Pass@1 | Pass@2 | +| --- | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | 186 | 57.80% | 65.05% | -## Select tasks by successful-attempt count +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. -The Parquet preserves the v2 columns (`instruct_successes`, `coder_successes` and their attempt counts) and adds `qwen38_successes`, `qwen38_attempts`, `qwen38_target_attempts` and `kept_from_v2`. Tasks already in v2 are not duplicated in this file. +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -```python -import pyarrow.parquet as pq +## Files and usage -# Newly evaluated tasks solved exactly once. -subset = pq.read_table("tasks.parquet", filters=[("qwen38_successes", "=", 1)]) -``` +`tasks.parquet` contains 121 newly kept tasks only. Tasks already kept in v2 are not duplicated. -[Selection metadata](metadata/selection.json) records membership for all 2,144 tasks, the v2 entry of each, and stage-1 evidence for the 186 evaluated tasks. [pass_counts.parquet](pass_counts.parquet) records the new evaluation; `retained_v3` identifies the incremental task set. +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. -## Traces +Success-count columns: `instruct_successes`, `coder_successes`, `qwen38_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. -`traces/` holds all 372 new attempts, successful and failed. Historical attempts remain in v2. `retained_v3` identifies tasks kept in this incremental release. +`archive.parquet` preserves excluded task payloads and reasons, including earlier exclusions where applicable. ```python -import json -import pyarrow.dataset as ds +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -traces = ds.dataset("traces", format="parquet", exclude_invalid_files=True) -new = traces.to_table(filter=ds.field("evaluation_version") == "v3") -conversation = json.loads(new["trajectory"][0].as_py()) +path = hf_hub_download( + "FWeindel/validated-tasks", + "endless_v3/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("qwen38_successes", "=", 1)]) ``` -All 372 new attempts have saved conversations and a verifier reward. 49 attempts hit the 300 s agent timeout; 11 of them still passed the verifier. No other errors were recorded. Traces are included without quality filtering. - -[Trace statistics](trace_stats/README.md) cover only the new evaluation. - -## Limitations - -- **Network:** isolation was requested, but every job fell back to host networking because the nodes had reached their network-namespace limit. Agents could have reached the network. -- **Instruction-path check:** for 3 of the 4 pilot tasks, the stage-1 path check searched folders that do not exist in these images and returned no result. All other stage-1 checks passed for them. -- **Stage-1 static checks:** 1 evaluated task failed `check-nproc`; it was not solved and stays excluded. -- **Repository-leak checks:** not yet applied to these tasks; see [v1](../endless_v1/README.md). - -## Runtime - -- **Versions:** Apptainer **1.5.4**, Python **3.12.14**, Harbor **0.8.1**. -- **Implementation:** [Marin Harbor `7faf878c14b6`](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) and [runtime patches](https://github.com/MLI-lab/OT-next-data/tree/main/harbor_patches). -- **Resources:** one H200 and 32 CPUs per job, 4 concurrent trials per GPU, 16 full-run shards with up to 8 running at once; each shard took 23 to 30 minutes. - -| Model | Mean time per attempt | Median time per attempt | Attempts measured | -| --- | ---: | ---: | ---: | -| Qwen/Qwen3.8-27B | 186.65 s | 189.12 s | 372 | - -`patch.py` rebuilds the incremental task file from the pinned pre-migration release and the published layout manifest. - -## Incremental release - -V3 contains only tasks newly solved after v2. It has no task-ID overlap with [v2](../endless_v2/README.md). This is a harder subset relative to the earlier evaluation, not a claim of intrinsic difficulty. New-evaluation traces include both solved and unsolved tasks so the reported pass rates remain reproducible. +Combine the 1,958 tasks in `endless_v2/tasks.parquet` with these 121 newly kept tasks for 2,079 unique tasks. Allow for different columns and retain each source file's image manifest when staging images. ```python -from huggingface_hub import hf_hub_download import pyarrow as pa -import pyarrow.parquet as pq def load(version): - path = hf_hub_download("FWeindel/validated-tasks", - f"endless_{version}/tasks.parquet", repo_type="dataset") + path = hf_hub_download( + "FWeindel/validated-tasks", + f"endless_{version}/tasks.parquet", + repo_type="dataset", + ) return pq.read_table(path) v2, v3 = load("v2"), load("v3") @@ -110,4 +84,14 @@ assert set(v2["path"].to_pylist()).isdisjoint(v3["path"].to_pylist()) combined = pa.concat_tables([v2, v3], promote_options="permissive") ``` -The example combines task rows; each source file retains its own image manifest. Use those source manifests when staging images. [Layout and provenance](metadata/layout.json) records the pinned baseline and task counts. Earlier cumulative v3 releases remain accessible through their HF commit history. +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. + +## Trace statistics + +Of 372 recorded attempts, 372 have saved conversations and 0 do not. 0 attempts have no raw verifier reward. Missing measurements remain missing. + +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. + +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. Historical attempts remain in v2. + +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/endless_v3/plots/readme-pass-at-k.png b/endless_v3/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..ac96f1d6cf2bf2dded21794dcaec79685f65fe92 --- /dev/null +++ b/endless_v3/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0826509fcb329b67639c345777cd6e8cbd8a1f60ba5e375b8e7aa911ca37d13 +size 37734 diff --git a/endless_v3/trace_stats/README.md b/endless_v3/trace_stats/README.md index 88b22ed8e6cd8862035c5e684d2e16164b579de5..a3e0df9a37bca379270ab0de98671a5a03e4a6fc 100644 --- a/endless_v3/trace_stats/README.md +++ b/endless_v3/trace_stats/README.md @@ -1,51 +1,53 @@ -# Trajectory statistics +# EndlessTerminals v3 trace statistics -372 attempts across 186 tasks of Endless Terminals v3, Qwen/Qwen3.8-27B evaluation, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3.8-27B | -| --- | ---: | -| Recorded attempts | 372 | -| Median agent turns | 4 | -| Median input tokens per attempt | 12,998 | -| Median output tokens per attempt | 9,777 | -| Median agent runtime (seconds) | 174 | -| Median verifier runtime (seconds) | 7 | +Of 372 recorded attempts, 372 have saved conversations and 0 do not. 0 attempts have no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. This page covers the new v3 evaluation; [v2 statistics](../../endless_v2/trace_stats/README.md) cover the historical traces. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. Historical attempts remain in v2. -![Trace measurements](plots/all-models-trace-distributions.png) +Token totals sum calls across an attempt and are not the context window. -![Trace outcomes](plots/all-models-trace-outcomes.png) +Measurements cover 186 tasks, including kept and unselected tasks, with each attempt counted once. -![Most frequent commands](plots/all-models-trace-commands.png) +## Evaluation settings -Commands are counted as submitted, not as successful executions. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | On | 2 | 1.0 | 0.95 | 20 | 32,768 | 8,192 | min_p=0; presence_penalty=0; repetition_penalty=1 | -Per-model plots: [Qwen/Qwen3.8-27B commands](plots/qwen3.8-27b-trace-commands.png), [Qwen/Qwen3.8-27B distributions](plots/qwen3.8-27b-trace-distributions.png), [Qwen/Qwen3.8-27B outcomes](plots/qwen3.8-27b-trace-outcomes.png). +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Data +**Time limits:** agent 300 seconds; verifier 300 seconds (all 372 saved configurations; limits vary by task where multiple values are shown). -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +**Output-limit evidence:** All 372 saved trial configurations contain an explicit 8,192-token request cap; the 37 saved pilot request payloads confirm it. Full-run HTTP payloads were not audited. -Attempts without a saved conversation: 0; without a verifier reward: 0. Their missing measurements stay empty. +Network isolation was requested, but the run fell back to host networking when network namespace creation failed; agents could have reached the network. +## Measurements -Evaluation budgets: see the [dataset evaluation setup](../README.md). Per-call context and output-limit enforcement are not established by token-usage statistics. +### Qwen/Qwen3.8-27B -## Evaluation settings +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 372 | 4 | 7 | +| Input tokens per attempt | 372 | 12,998 | 30,223 | +| Output tokens per attempt | 372 | 9,777 | 16,073 | +| Largest recorded step, prompt plus output tokens | 372 | 9,086 | 13,907 | +| Agent runtime (seconds) | 372 | 174.45 | 300 | +| Verifier runtime (seconds) | 372 | 7.29 | 8.59 | +| Whole-attempt runtime (seconds) | 372 | 189.12 | 317.02 | -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3.8-27B | 32768 | 8192 | `llm_call_kwargs.max_tokens` = 8192 in all 372 saved trial configs; 37 pilot request payloads checked | on | {"min_p": 0.0, "presence_penalty": 0.0, "repetition_penalty": 1.0, "temperature": 1.0, "top_k": 20, "top_p": 0.95} | +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -Configured limits are not measurements of token usage. Agent and verifier time limits are 300 s each for every evaluated task. Unlike the v2 runs, this run forwards the output cap to each request; the earlier attempt at this run (job 1065977) did not and was cancelled before it produced results. +## Plots -Settings and evidence: [audit](../metadata/evaluation-settings-audit.json). +![Trace measurements](plots/all-models-trace-distributions.png) -This folder reports the new evaluation only. Task membership is incremental: tasks already in v2 are not duplicated in v3. Historical conversations remain in v2. See the [release overview](../README.md) for counts and combination instructions. +![Trace outcomes](plots/all-models-trace-outcomes.png) + +![Most frequent commands](plots/all-models-trace-commands.png) diff --git a/facet_v2/README.md b/facet_v2/README.md index 525b3b32f1fac57ae1fcba7bdf2fb0040b2f9663..582448e92341e4c301fe2713b6de16866e7866d8 100644 --- a/facet_v2/README.md +++ b/facet_v2/README.md @@ -1,80 +1,77 @@ -# FACET v2: Qwen/Qwen3-Coder-30B-A3B-Instruct pass@4 release +# FACET v2 -- **Original tasks:** 6,020 -- **Kept:** 780 -- **Changed task payloads:** 0 -- **Archived:** 5,240 (5,145 with no observed success, plus 95 prior v1 exclusions) -- **Pending:** 0 -- **Evaluated:** all 5,925 v1 tasks -- **Recorded attempts:** 23,700 -- **Saved conversations:** 23,700; no missing conversations -- **Successful attempts:** 1,591 +## Dataset stats -`tasks.parquet` contains the 780 tasks solved at least once in four recorded attempts. `archive.parquet` preserves unsuccessful tasks and prior exclusions with their reasons. All task payloads are unchanged from [facet_v1](../facet_v1), pinned at `3abfc63260ef8ab59ac61274450310f650585bc9`. This completes evaluation coverage of the v1 tasks. +- **Tasks:** 5,925 in v1; 780 kept in v2. +- **Teacher evaluation:** Qwen/Qwen3-Coder-30B-A3B-Instruct (thinking off), 4 attempts per task (pass@4). +- **Coverage:** 5,925/5,925 tasks evaluated; 23,700 evaluation attempts. +- **Saved conversations:** 23,700 of 23,700 published attempt records, including successful and unsuccessful attempts. -## Evaluation setup and coverage - -Qwen/Qwen3-Coder-30B-A3B-Instruct with Terminus-2, four attempts per task. The completed pilot and 247 full-run shards supply these results, including the recovered shards. Every task has four recorded attempts, each counted once. +We keep tasks with at least one verifier-successful attempt from any evaluated model. ## Evaluation settings -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-Coder-30B-A3B-Instruct | Terminus-2 | Off | 4 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | repetition_penalty=1.05 | -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-Coder-30B-A3B-Instruct | 32768 | 8192 | top-level only; explicit request cap not forwarded by Terminus | off | {"repetition_penalty": 1.05, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +**Time limits:** agent 120, 180, 240, 1,800 seconds; verifier 120 seconds (all saved configurations; limits vary by task where multiple values are shown). Agent limits are 120 seconds for 23,676 attempts, 180 for 12, 240 for 8 and 1,800 for 4. +**Output-limit evidence:** All 23,700 saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. -Both frozen contracts use the settings above. The intended 8,192-token output cap was stored at the top level of agent kwargs and was **not forwarded** as an explicit request cap by the historical Terminus integration. All 23,700 saved configurations were checked. The actual fallback cap is unknown. This is a per-call limit, not the sum of output tokens across a trajectory. The pinned weights revision is `b2cff646eb4bb1d68355c01b18ae02e7cf42d120`. +## Results -| Agent time limit | Tasks | Attempts | Verifier time limit | -| --- | ---: | ---: | --- | -| 120 seconds | 5,919 | 23,676 | 120 seconds | -| 180 seconds | 3 | 12 | 120 seconds | -| 240 seconds | 2 | 8 | 120 seconds | -| 1,800 seconds | 1 | 4 | 120 seconds | +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. -[Settings audit](metadata/evaluation-settings-audit.json), [frozen contracts](metadata/full-contract.json), and [selection and raw attempt rewards](metadata/selection.json) preserve the evidence. +| Successful attempts | Qwen/Qwen3-Coder-30B-A3B-Instruct (5,925 tasks) | +| --- | ---: | +| 0/4 | 5,145 | +| 1/4 | 348 | +| 2/4 | 176 | +| 3/4 | 133 | +| 4/4 | 123 | -This release adds teacher results and saved stage-7 evidence; it does not establish that the older v1 validation satisfies today's complete default profile. In particular, outstanding Git-history and timestamp leak-check coverage is not claimed as passed. +| Model | Evaluated tasks | Pass@1 | Pass@2 | Pass@4 | +| --- | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-Coder-30B-A3B-Instruct | 5,925 | 6.71% | 9.73% | 13.16% | -## Results across all evaluated tasks +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. -Rates use all 5,925 tasks with four recorded attempts. Pass@1 through pass@3 are estimates from these four samples; pass@4 is the fraction solved at least once. +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -| Metric | Rate | -| --- | ---: | -| Pass@1 | 6.71% | -| Pass@2 | 9.73% | -| Pass@3 | 11.70% | -| Pass@4 | 13.16% | +## Files and usage -| Successful attempts out of four | Tasks | -| --- | ---: | -| 0 | 5,145 | -| 1 | 348 | -| 2 | 176 | -| 3 | 133 | -| 4 | 123 | +`tasks.parquet` contains 780 kept tasks. + +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. + +Success-count columns: `coder_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. + +`archive.parquet` preserves excluded task payloads and reasons, including earlier exclusions where applicable. + +```python +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -Observed reward groups: 5,145 tasks always scored zero, 123 always succeeded, and 657 varied across attempts. No task had constant partial credit. These groups describe the recorded samples. +path = hf_hub_download( + "FWeindel/validated-tasks", + "facet_v2/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("coder_successes", "=", 1)]) +``` -[Pass-rate plot](plots/pass-rates.pdf) and [success-count plot](plots/solved-counts.pdf). +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. -## Trajectories, runtime and tokens +## Trace statistics -All 23,700 recorded conversations are included in [traces/coder.parquet](traces/coder.parquet), including unsuccessful attempts. Availability is not a claim that every conversation passes every quality check. Five recorded attempts have no raw verifier reward; the historical terminal-error policy counts these as unsolved while preserving their null raw reward. There are no missing attempts. The reports record 8,514 agent timeouts, 3 runtime errors and 2 phase-cancellation errors. +Of 23,700 recorded attempts, 23,700 have saved conversations and 0 do not. 5 attempts have no raw verifier reward. Missing measurements remain missing. -The median attempt has 31 agent turns, 355,277 cumulative input tokens, 6,474 output tokens, 82 seconds of agent runtime and 7 seconds of verifier runtime. Input tokens include repeated context across calls. Medians use available measurements. See [trajectory statistics and distributions](trace_stats/README.md) for sample counts, outcomes, timeouts, per-task and per-attempt measurements, and token accounting across the complete evaluation. No API cost or GPU-utilization estimate is inferred from these timings. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Files and reproduction +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -- `tasks.parquet`: 780 kept tasks, with pinned image references. -- `archive.parquet`: 5,240 excluded tasks, preserving archive reasons. -- `pending.parquet`: an empty table, retained for compatibility; no tasks remain pending. -- `traces/`: all recorded conversations, raw trial results and their manifest. -- `trace_stats/`: statistics and plots across all 23,700 attempts and 5,925 tasks. -- `metadata/`: frozen contracts, task hashes, shard provenance, complete stage-6 outcomes, recovery stage-7 reports and selection evidence. -- [patch.py](patch.py): reproduce task membership and payloads from the pinned v1 release and saved outcomes. +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/facet_v2/plots/readme-pass-at-k.png b/facet_v2/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..f99a0523035e7f1881fa3396a9c3a71cf9296347 --- /dev/null +++ b/facet_v2/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b58d0839c5c1b0e5114ece0ff318766d298bdd537620ee08313cc708a056f4d +size 38779 diff --git a/facet_v2/trace_stats/README.md b/facet_v2/trace_stats/README.md index 2bf7f91d8acf95d9347ad132fbd6bcafc24042b1..e3eff60c9d14968edfc1ab6e66ce7190c2bdc0c4 100644 --- a/facet_v2/trace_stats/README.md +++ b/facet_v2/trace_stats/README.md @@ -1,64 +1,55 @@ -# Trajectory statistics +# FACET v2 trace statistics -23,700 attempts across 5,925 tasks of facet_v2, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3-Coder-30B-A3B-Instruct | -| --- | ---: | -| Recorded attempts | 23,700 | -| Median agent turns | 31 | -| Median input tokens per attempt | 355,277 | -| Median output tokens per attempt | 6,474 | -| Median agent runtime (seconds) | 82 | -| Median verifier runtime (seconds) | 7 | +Of 23,700 recorded attempts, 23,700 have saved conversations and 0 do not. 5 attempts have no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots - -![Trace measurements](plots/all-models-trace-distributions.png) - -![Trace outcomes](plots/all-models-trace-outcomes.png) - -![Most frequent commands](plots/all-models-trace-commands.png) +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -Commands are counted as submitted, not as successful executions. +Token totals sum calls across an attempt and are not the context window. -Per-model plots: [Qwen/Qwen3-Coder-30B-A3B-Instruct commands](plots/qwen3-coder-30b-a3b-instruct-trace-commands.png), [Qwen/Qwen3-Coder-30B-A3B-Instruct distributions](plots/qwen3-coder-30b-a3b-instruct-trace-distributions.png), [Qwen/Qwen3-Coder-30B-A3B-Instruct outcomes](plots/qwen3-coder-30b-a3b-instruct-trace-outcomes.png). +Measurements cover 5,925 tasks, including kept and unselected tasks, with each attempt counted once. -## Data - -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +## Evaluation settings -Attempts without a saved conversation: 0; without a verifier reward: 5. Their missing measurements stay empty. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-Coder-30B-A3B-Instruct | Terminus-2 | Off | 4 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | repetition_penalty=1.05 | +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Evaluation settings +**Time limits:** agent 120, 180, 240, 1,800 seconds; verifier 120 seconds (all saved configurations; limits vary by task where multiple values are shown). Agent limits are 120 seconds for 23,676 attempts, 180 for 12, 240 for 8 and 1,800 for 4. -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +**Output-limit evidence:** All 23,700 saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-Coder-30B-A3B-Instruct | 32768 | 8192 | top-level only; explicit request cap not forwarded by Terminus | off | {"repetition_penalty": 1.05, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | -| Qwen/Qwen3-Coder-30B-A3B-Instruct | 32768 | 8192 | top-level only; explicit request cap not forwarded by Terminus | off | {"repetition_penalty": 1.05, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +## Measurements -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +### Qwen/Qwen3-Coder-30B-A3B-Instruct +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 23,700 | 31 | 61 | +| Input tokens per attempt | 23,700 | 355,277 | 880,082 | +| Output tokens per attempt | 23,700 | 6,474 | 10,799 | +| Largest recorded step, prompt plus output tokens | 23,697 | 20,765 | 24,794 | +| Agent runtime (seconds) | 23,700 | 82.34 | 120 | +| Verifier runtime (seconds) | 23,695 | 7.36 | 9.34 | +| Whole-attempt runtime (seconds) | 23,700 | 98.93 | 140.93 | -## Token usage and summarization +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -All published traces are included, covering kept and excluded tasks and every verifier outcome. One trace means one task attempt. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -![Output, summarization, and context distributions](plots/all-models-tokens-by-outcome.png) +## Plots -- **Output per recorded agent step:** successful, graded unsuccessful, and ungraded attempts shown separately. Each step contributes once. -- **Summarizations per trace:** one combined distribution across all traces with a recorded count, including zero. -- **Peak context per trace:** the largest recorded step prompt + output count in each trace, separated into traces with and without summarization. Unknown summary counts are kept separate when present. Each trace contributes once, regardless of its length or outcome. +![Trace measurements](plots/all-models-trace-distributions.png) -[Detailed statistics](data/tokens-by-outcome.json) retain sample counts and missing measurements. Missing summary counts are not zero. ATIF steps can aggregate multiple requests, so recorded context does not establish an exact per-call context limit. +![Trace outcomes](plots/all-models-trace-outcomes.png) +![Most frequent commands](plots/all-models-trace-commands.png) -## Recorded time limits +![Output, summarization and recorded context](plots/all-models-tokens-by-outcome.png) -The agent limit was 120 seconds for 23,676 attempts (5,919 tasks), 180 seconds for 12 attempts (3 tasks), 240 seconds for 8 attempts (2 tasks), and 1,800 seconds for 4 attempts (1 task). The verifier limit was 120 seconds throughout. The [settings audit](../metadata/evaluation-settings-audit.json) records all 23,700 configurations and their counts. The intended 8,192-token response cap was not forwarded as an explicit request cap; the actual fallback cap is unknown. +Output per recorded agent step is separated by verifier outcome. Summarizations are counted per trace, including zero when recorded. Peak recorded context is shown per trace with and without summarization; unknown counts stay separate. diff --git a/fineenvs_v2/README.md b/fineenvs_v2/README.md index f4cba1d76e14314c6db5e346a54712225c8dc8c1..e925c5db9823500c0dd86f46640d2d445748367b 100644 --- a/fineenvs_v2/README.md +++ b/fineenvs_v2/README.md @@ -1,74 +1,76 @@ # FineEnvs v2 -- **Original tasks:** 5,000 -- **Kept:** 4,001 -- **Changed task payloads from v1:** 0 -- **Archived:** 999 with no observed success in four attempts -- **Pending evaluations:** 0 -- **Evaluated tasks:** 5,000 -- **Recorded attempts:** 20,000 -- **Saved conversations:** 20,000 -- **Verifier-graded attempts:** 19,973 -- **Successful attempts:** 13,438 -- **Images:** 1, reused from the pinned v1 release +## Dataset stats -FineEnvs asks agents to inspect packaged tabular datasets, compute answers with Python data-analysis tools, and write them to `/workdir/answer.txt`. Tasks cover statistics, correlations, filtering, aggregation and model evaluation. +- **Tasks:** 5,000 in v1; 4,001 kept in v2. +- **Teacher evaluation:** Qwen/Qwen3-30B-A3B-Instruct-2507 (thinking off), 4 attempts per task (pass@4). +- **Coverage:** 5,000/5,000 tasks evaluated; 20,000 evaluation attempts. +- **Saved conversations:** 20,000 of 20,000 published attempt records, including successful and unsuccessful attempts. -This release keeps tasks solved at least once by Qwen/Qwen3-30B-A3B-Instruct-2507. Task payloads are unchanged from [fineenvs_v1](../fineenvs_v1), revision `7146b8548f19c3474827162969b270ddcae76b5f`, which removed the unused `HF_TOKEN` requirement. `archive.parquet` preserves all 999 other tasks with their selection reason. All recorded attempts, including unsuccessful attempts and attempts for archived tasks, are included in `traces/`. - -## Evaluation setup and coverage - -Qwen/Qwen3-30B-A3B-Instruct-2507 (thinking disabled) ran with Terminus-2 for four attempts per task on Helma. All 100 shards completed on October 22 and 23, 2026. Stages 6 and 7 produced complete reports for all 5,000 tasks. This release adds teacher evaluation and trace statistics; it does not rerun or recertify the v1 validation checks against the current default profile. +We keep tasks with at least one verifier-successful attempt from any evaluated model. ## Evaluation settings -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 4 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling Agent limit (seconds) | Tasks | -| --- | ---: | ---: | --- | --- | --- ---: | ---: | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | present in saved request kwargs; HTTP payload not audited | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} 600 | 4,938 | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | present in saved request kwargs; HTTP payload not audited | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} 900 | 62 | +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -Configured limits are not measurements of token usage. Unknown historical settings are not filled from current model defaults. +**Time limits:** agent 600, 900 seconds; verifier 120 seconds (all saved configurations; limits vary by task where multiple values are shown). Agent limits are 600 seconds for 19,752 attempts and 900 for 248. No separate total trial timeout is configured. +**Output-limit evidence:** All 20,000 saved configurations contain an explicit 8,192-token request cap. HTTP payloads and server enforcement were not audited. -Saved configurations were checked for all 20,000 attempts. The 8,192-token output cap is present in `llm_call_kwargs.max_tokens`; HTTP payloads and server enforcement were not independently audited. Context is 32,768 tokens per call. Sampling is temperature 0.7, top-p 0.8, top-k 20 and min-p 0. Agent limits are 600 seconds for 4,938 tasks and 900 seconds for 62 tasks; verifier limits are 120 seconds, and environment setup limits are 600 seconds. There is no separate total trial timeout in the saved configurations. See [settings evidence](metadata/evaluation-settings-audit.json), [task resources](metadata/task-resources.json) and [frozen contract](metadata/contract.json). +## Results -## Results over all 5,000 evaluated tasks +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. -| Metric | Rate | +| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 (5,000 tasks) | | --- | ---: | -| Pass@1 | 67.19% | -| Pass@2 | 75.31% | -| Pass@4 | 80.02% | +| 0/4 | 999 | +| 1/4 | 334 | +| 2/4 | 410 | +| 3/4 | 744 | +| 4/4 | 2,513 | -These rates describe the full v1 evaluation cohort, before selecting the 4,001 kept tasks. Pass@k uses the usual sampling-without-replacement estimator from four recorded attempts. +| Model | Evaluated tasks | Pass@1 | Pass@2 | Pass@4 | +| --- | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | 5,000 | 67.19% | 75.31% | 80.02% | -| Successful attempts out of four | Tasks | -| --- | ---: | -| 0 | 999 | -| 1 | 334 | -| 2 | 410 | -| 3 | 744 | -| 4 | 2,513 | +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. + +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) + +## Files and usage + +`tasks.parquet` contains 4,001 kept tasks. + +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. + +`archive.parquet` preserves excluded task payloads and reasons, including earlier exclusions where applicable. + +```python +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq + +path = hf_hub_download( + "FWeindel/validated-tasks", + "fineenvs_v2/tasks.parquet", + repo_type="dataset", +) +tasks = pq.read_table(path) +``` -Observed variability: 2,513 tasks were always solved, 1,488 were sometimes solved, and 999 had zero successes. These describe the sampled attempts, not intrinsic task difficulty. +Per-task success counts are stored in `metadata/selection.json`; this task file preserves the v1 schema. -![Pass rates](plots/pass-rates.png) -![Success counts](plots/solved-counts.png) +This teacher evaluation does not rerun or recertify v1 validation against the current default profile. -## Trajectory quality, failures and runtime +## Trace statistics -27 attempts across 13 tasks have no verifier reward. The recorded `terminal-errors-zero-v2` policy counts them as unsolved for pass@k, while the raw missing rewards remain null in the traces and selection evidence. Errors include terminal protocol/runtime failures and one verifier setup failure. Eight of the 999 archived tasks include at least one such ungraded attempt, so their selection means no observed success, not four verifier-confirmed failures. These are not successful checks. A saved conversation is not a claim that it passed every quality check; no additional conversation filtering is applied here. +Of 20,000 recorded attempts, 20,000 have saved conversations and 0 do not. 27 attempts have no raw verifier reward. Missing measurements remain missing. -The median attempt has 4 agent turns, 7,904 cumulative input tokens, 942 output tokens, 10 seconds of agent runtime and 8 seconds of verifier runtime. Input tokens count repeated context across calls. These are measured attempt distributions, not batch elapsed times. See [trajectory statistics](trace_stats/README.md) for distributions, termination reasons, errors and measurement coverage. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Files and reproduction +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -- [tasks.parquet](tasks.parquet): 4,001 kept task packages and the pinned v1 image references. -- [archive.parquet](archive.parquet): 999 task packages with no observed success. -- [pass_counts.parquet](pass_counts.parquet): counts and pass@k for all 5,000 evaluated tasks, distinguishing recorded from verifier-graded attempts. -- [traces/instruct.parquet](traces/instruct.parquet): all 20,000 attempts with conversations, trial records, raw verifier rewards and selection membership. -- [metadata/selection.json](metadata/selection.json): per-task selection and recorded outcomes. -- [Stage 6](metadata/stage-6-merged.json) and [stage 7](metadata/stage-7-merged.json): original merged reports. -- [patch.py](patch.py): reproduce task membership from the pinned v1 source and selection evidence. +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/fineenvs_v2/plots/readme-pass-at-k.png b/fineenvs_v2/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..07dfc69809bfc43c894257a99efab5ea110a1832 --- /dev/null +++ b/fineenvs_v2/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4eaedd7235bf87a232515a656e216cc59e02f92d08139c4e5a62b45e428e88e +size 39668 diff --git a/fineenvs_v2/trace_stats/README.md b/fineenvs_v2/trace_stats/README.md index bb5dbf9347e67d264ba90e572e3172f3f09b54d2..9d5f25b1442aa12c3ed66fe2c7f3dd77922d57f5 100644 --- a/fineenvs_v2/trace_stats/README.md +++ b/fineenvs_v2/trace_stats/README.md @@ -1,61 +1,55 @@ -# Trajectory statistics +# FineEnvs v2 trace statistics -20,000 attempts across 5,000 tasks of fineenvs_v2, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3-30B-A3B-Instruct-2507 | -| --- | ---: | -| Recorded attempts | 20,000 | -| Median agent turns | 4 | -| Median input tokens per attempt | 7,904 | -| Median output tokens per attempt | 942 | -| Median agent runtime (seconds) | 10 | -| Median verifier runtime (seconds) | 8 | +Of 20,000 recorded attempts, 20,000 have saved conversations and 0 do not. 27 attempts have no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots - -![Trace measurements](plots/all-models-trace-distributions.png) +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -![Trace outcomes](plots/all-models-trace-outcomes.png) - -![Most frequent commands](plots/all-models-trace-commands.png) +Token totals sum calls across an attempt and are not the context window. -Commands are counted as submitted, not as successful executions. +Measurements cover 5,000 tasks, including kept and unselected tasks, with each attempt counted once. -Per-model plots: [Qwen/Qwen3-30B-A3B-Instruct-2507 commands](plots/qwen3-30b-a3b-instruct-2507-trace-commands.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 distributions](plots/qwen3-30b-a3b-instruct-2507-trace-distributions.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 outcomes](plots/qwen3-30b-a3b-instruct-2507-trace-outcomes.png). +## Evaluation settings -## Data +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 4 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -Attempts without a saved conversation: 0; without a verifier reward: 27. Their missing measurements stay empty. +**Time limits:** agent 600, 900 seconds; verifier 120 seconds (all saved configurations; limits vary by task where multiple values are shown). Agent limits are 600 seconds for 19,752 attempts and 900 for 248. No separate total trial timeout is configured. +**Output-limit evidence:** All 20,000 saved configurations contain an explicit 8,192-token request cap. HTTP payloads and server enforcement were not audited. -## Evaluation settings +## Measurements -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +### Qwen/Qwen3-30B-A3B-Instruct-2507 -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling Agent limit (seconds) | Tasks | -| --- | ---: | ---: | --- | --- | --- ---: | ---: | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | present in saved request kwargs; HTTP payload not audited | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} 600 | 4,938 | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | present in saved request kwargs; HTTP payload not audited | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} 900 | 62 | +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 20,000 | 4 | 8 | +| Input tokens per attempt | 20,000 | 7,904 | 24,714 | +| Output tokens per attempt | 20,000 | 942 | 2,161 | +| Largest recorded step, prompt plus output tokens | 19,999 | 2,692 | 5,504 | +| Agent runtime (seconds) | 20,000 | 9.68 | 25.82 | +| Verifier runtime (seconds) | 19,974 | 7.89 | 8.65 | +| Whole-attempt runtime (seconds) | 20,000 | 27.08 | 47.82 | -Configured limits are not measurements of token usage. Unknown historical settings are not filled from current model defaults. +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -## Token usage and summarization +## Plots -All published traces are included, covering kept and excluded tasks and every verifier outcome. One trace means one task attempt. +![Trace measurements](plots/all-models-trace-distributions.png) -![Output, summarization, and context distributions](plots/all-models-tokens-by-outcome.png) +![Trace outcomes](plots/all-models-trace-outcomes.png) -- **Output per recorded agent step:** successful, graded unsuccessful, and ungraded attempts shown separately. Each step contributes once. -- **Summarizations per trace:** one combined distribution across all traces with a recorded count, including zero. -- **Peak context per trace:** the largest recorded step prompt + output count in each trace, separated into traces with and without summarization. Unknown summary counts are kept separate when present. Each trace contributes once, regardless of its length or outcome. +![Most frequent commands](plots/all-models-trace-commands.png) -[Detailed statistics](data/tokens-by-outcome.json) retain sample counts and missing measurements. Missing summary counts are not zero. ATIF steps can aggregate multiple requests, so recorded context does not establish an exact per-call context limit. +![Output, summarization and recorded context](plots/all-models-tokens-by-outcome.png) -Agent time limits are 600 or 900 seconds by task, verifier limits are 120 seconds, and environment setup limits are 600 seconds. No separate total trial timeout is configured. [Task resource evidence](../metadata/task-resources.json). +Output per recorded agent step is separated by verifier outcome. Summarizations are counted per trace, including zero when recorded. Peak recorded context is shown per trace with and without summarization; unknown counts stay separate. diff --git a/nl2bash_v2/README.md b/nl2bash_v2/README.md index 9052d0bbdd37a476ce96245d79fc666eacbf0020..ea2ecfeb77adaa2d28f332b26610446be37d81e5 100644 --- a/nl2bash_v2/README.md +++ b/nl2bash_v2/README.md @@ -2,127 +2,76 @@ ## Dataset stats -- **Original tasks:** 1,497 -- **Kept:** 938 -- **Changed from v1:** 0 task payloads -- **Archived:** 559 (487 with no Qwen/Qwen3-Coder-30B-A3B-Instruct success and 72 earlier v1 exclusions) -- **Pending attempts:** 0 -- **Evaluated tasks:** 1,425 -- **Recorded attempts:** 5,700 -- **Saved conversations:** 5,700 -- **Successful attempts:** 2,911 -- **Images:** 1 before, 1 now - -Based on [nl2bash_v1](https://huggingface.co/datasets/FWeindel/validated-tasks/tree/3abfc63260ef8ab59ac61274450310f650585bc9/nl2bash_v1) at revision `3abfc63260ef8ab59ac61274450310f650585bc9`. +- **Tasks:** 1,425 in v1; 938 kept in v2. +- **Teacher evaluation:** Qwen/Qwen3-Coder-30B-A3B-Instruct (thinking off), 4 attempts per task (pass@4). +- **Coverage:** 1,425/1,425 tasks evaluated; 5,700 evaluation attempts. +- **Saved conversations:** 5,700 of 5,700 published attempt records, including successful and unsuccessful attempts. -## Evaluation settings - -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. - -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-Coder-30B-A3B-Instruct | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"repetition_penalty": 1.05, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | - -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. - -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. - -The table reports settings found in **1 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. - -## Dataset characteristics and changes from v1 - -[NL2Bash](https://arxiv.org/abs/1802.08979) tasks from [TaskTrove](https://huggingface.co/datasets/open-thoughts/TaskTrove) `nl2bash` at revision `946884702046be1a7dcea2638186ad6b0d2ea103`. +We keep tasks with at least one verifier-successful attempt from any evaluated model. -
Domains - -**File Operations:** locate, filter, count, sort or copy files in a seeded `/workspace` with find, grep, xargs, du, sort and cp, then capture the command's output in `/output/command_capture.txt`. **System Administration:** compose shell one-liners with coreutils, variables, redirection and pipelines in an Ubuntu container. - -
-
Capabilities - -**Specification Adherence:** each request stacks precise details (depth limit, case-insensitivity, regular files only, numeric sort, capturing both stdout and stderr), and the verifier compares the output lines with the reference command's output, so dropping one detail fails the task. Labels follow [CLI-Universe](https://arxiv.org/abs/2606.22883v1). - -
- -Intended benchmark transfer: **Terminal-Bench 2.0, OpenThoughts-TBLite 2.0 and Harbor-Index**, through translating requests into correct shell pipelines in a Harbor-style container. - -Kept 938 tasks solved at least once in four attempts. Task payloads and the pinned v1 image are unchanged. Archived 487 tasks with zero observed successes under this evaluation, preserving their payloads and attempt evidence. Earlier archive reasons are preserved for all 72 v1 exclusions. +## Evaluation settings -The v1 release documents build/setup, reference-solution, no-op and rule-based validation. Its Git-history and timestamp leak checks were not run. This release filters by teacher results and does not establish completion of those outstanding validation checks. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-Coder-30B-A3B-Instruct | Terminus-2 | Off | 4 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | repetition_penalty=1.05 | -## Evaluation setup and coverage +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -| Setting | Value | -| --- | --- | -| Teacher | Qwen/Qwen3-Coder-30B-A3B-Instruct | -| Agent | Terminus 2 | -| Coverage | 1,425 tasks, four attempts each, 5,700 total | -| Sampling | Temperature 0.7; top-p 0.8; top-k 20; repetition penalty 1.05 | -| Token budgets | 32,768 context; intended 8,192 output tokens per call (see enforcement caveat) | -| Serving | One H200 per shard; eight concurrent trials; 32 CPUs, including four serving CPUs | -| Evidence | 61 complete stage-6 reports and 61 complete stage-7 reports, including the disjoint eight-task pilot | +**Time limits:** agent 600 seconds; verifier 600 seconds (sampled configurations; limits vary by task where multiple values are shown). -The 938 kept tasks each declare a 600-second agent timeout and a 600-second verifier timeout. [Frozen contracts and provenance](metadata/provenance.json) record model revision, execution settings, task hashes and input evidence. Only the replacement pilot and full-array results are included. +**Output-limit evidence:** The 1 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -## Pass rates and success distribution +## Results -All rates cover the full 1,425-task evaluated set, before filtering. Higher-k values use the saved four-attempt samples. +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. -| Metric | Rate | +| Successful attempts | Qwen/Qwen3-Coder-30B-A3B-Instruct (1,425 tasks) | | --- | ---: | -| pass@1 | 51.07% | -| pass@2 | 59.79% | -| pass@3 | 63.65% | -| pass@4 | 65.82% | - -| Successful attempts out of four | Tasks | -| ---: | ---: | -| 0 | 487 | -| 1 | 124 | -| 2 | 144 | -| 3 | 181 | -| 4 | 489 | +| 0/4 | 487 | +| 1/4 | 124 | +| 2/4 | 144 | +| 3/4 | 181 | +| 4/4 | 489 | -Observed reward groups: 489 always solved, 487 always zero, and 449 varying. These describe the recorded samples, not intrinsic task difficulty. +| Model | Evaluated tasks | Pass@1 | Pass@2 | Pass@4 | +| --- | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-Coder-30B-A3B-Instruct | 1,425 | 51.07% | 59.79% | 65.82% | -![Pass rates](plots/pass-rates.png) +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. -![Successful attempts per task](plots/solved-counts.png) +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -## Trajectories and trace endings +## Files and usage -All 5,700 saved conversations are included in [traces/coder.parquet](traces/coder.parquet), with original result records, task and attempt IDs, raw verifier rewards, effective rewards and the `retained_v2` flag. No conversations were filtered out of this evidence export. Of these, 2,911 attempts received verifier reward 1. +`tasks.parquet` contains 938 kept tasks. -There are 5,699 verifier grades and one ungraded terminal `TmuxSessionEndedError`. Following the recorded `terminal-errors-zero-v2` policy, its effective reward is zero; its raw `verifier_reward` stays null. Its task has no success in the other three attempts and is archived. No absent attempts were filled in. +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. -Trace endings: 4,249 task-complete and 1,451 tmux-session-closed. Session closure and grading are separate: 1,450 of those closures still have a verifier grade. Stage 7 completed for all tasks; presence of a conversation is not an additional training-quality acceptance claim. +Success-count columns: `coder_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. -![Trace outcomes](trace_stats/plots/all-models-trace-outcomes.png) +`archive.parquet` preserves excluded task payloads and reasons, including earlier exclusions where applicable. -## Runtime, tokens and resources +```python +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -| Measurement | Attempts measured | Median | Maximum | -| --- | ---: | ---: | ---: | -| Agent turns | 5,700 | 11 | 497 | -| Input tokens per attempt | 5,700 | 25,332 | 6,915,919 | -| Output tokens per attempt | 5,700 | 1,529 | 54,555 | -| Largest recorded turn, input plus output tokens | 5,700 | 3,625 | 35,245 | -| Agent time, seconds | 5,700 | 17.31 | 554.40 | -| Verifier time, seconds | 5,699 | 5.87 | 12.47 | -| Trial time, seconds | 5,700 | 31.71 | 568.58 | +path = hf_hub_download( + "FWeindel/validated-tasks", + "nl2bash_v2/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("coder_successes", "=", 1)]) +``` -Input and output totals sum model calls within each attempt. Recorded input-plus-output turn size is a trace measurement, not a configured serving context limit. Per-task peak memory is not reported here. +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. -![Trace distributions](trace_stats/plots/all-models-trace-distributions.png) +## Trace statistics -See [trajectory statistics](trace_stats/README.md) for distributions, command counts, error categories and machine-readable per-attempt metrics. +Of 5,700 recorded attempts, 5,700 have saved conversations and 0 do not. 1 attempt has no raw verifier reward. Missing measurements remain missing. -## Files and selection +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -- [tasks.parquet](tasks.parquet): the 938 tasks solved at least once, with `coder_successes`, `coder_attempts` and `coder_graded_attempts` columns. -- [archive.parquet](archive.parquet): 559 excluded tasks, preserving earlier exclusion reasons. The 72 earlier exclusions have null Qwen/Qwen3-Coder-30B-A3B-Instruct counts because they were not evaluated. -- [selection evidence](metadata/selection.json): per-task success counts and raw attempt grades. -- [trace manifest](traces/manifest.json): trace coverage, reward policy and file checksum. -- [patch.py](patch.py): reproduces the kept task file from the pinned v1 source and saved selection. +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -The task Parquet preserves pinned references to the existing v1 image bundle; no image rebuild is needed for this selection. +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/nl2bash_v2/plots/readme-pass-at-k.png b/nl2bash_v2/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..c8a41bf170950726091352c99d1816fa9c1bc99a --- /dev/null +++ b/nl2bash_v2/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0c49ff4915efa4db75b7f8da3859c091cd0912d8b77f95edd9617f7c823869b +size 41289 diff --git a/nl2bash_v2/trace_stats/README.md b/nl2bash_v2/trace_stats/README.md index 39388a22951e883e5f15c5296f9533abe4d8e876..99cdfb83b68d62fba02744bf0791f2373c472d00 100644 --- a/nl2bash_v2/trace_stats/README.md +++ b/nl2bash_v2/trace_stats/README.md @@ -1,48 +1,51 @@ -# Trajectory statistics +# NL2Bash v2 trace statistics -5,700 attempts across 1,425 tasks of publication, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3-Coder-30B-A3B-Instruct | -| --- | ---: | -| Recorded attempts | 5,700 | -| Median agent turns | 11 | -| Median input tokens per attempt | 25,332 | -| Median output tokens per attempt | 1,529 | -| Median agent runtime (seconds) | 17 | -| Median verifier runtime (seconds) | 6 | +Of 5,700 recorded attempts, 5,700 have saved conversations and 0 do not. 1 attempt has no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -![Trace measurements](plots/all-models-trace-distributions.png) +Token totals sum calls across an attempt and are not the context window. -![Trace outcomes](plots/all-models-trace-outcomes.png) +Measurements cover 1,425 tasks, including kept and unselected tasks, with each attempt counted once. -![Most frequent commands](plots/all-models-trace-commands.png) +## Evaluation settings -Commands are counted as submitted, not as successful executions. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-Coder-30B-A3B-Instruct | Terminus-2 | Off | 4 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | repetition_penalty=1.05 | -Per-model plots: [Qwen/Qwen3-Coder-30B-A3B-Instruct commands](plots/qwen3-coder-30b-a3b-instruct-trace-commands.png), [Qwen/Qwen3-Coder-30B-A3B-Instruct distributions](plots/qwen3-coder-30b-a3b-instruct-trace-distributions.png), [Qwen/Qwen3-Coder-30B-A3B-Instruct outcomes](plots/qwen3-coder-30b-a3b-instruct-trace-outcomes.png). +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Data +**Time limits:** agent 600 seconds; verifier 600 seconds (sampled configurations; limits vary by task where multiple values are shown). -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +**Output-limit evidence:** The 1 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -Attempts without a saved conversation: 0; without a verifier reward: 1. Their missing measurements stay empty. +## Measurements -## Evaluation settings +### Qwen/Qwen3-Coder-30B-A3B-Instruct -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 5,700 | 11 | 18 | +| Input tokens per attempt | 5,700 | 25,332 | 60,036 | +| Output tokens per attempt | 5,700 | 1,529 | 2,661 | +| Largest recorded step, prompt plus output tokens | 5,700 | 3,625 | 5,707 | +| Agent runtime (seconds) | 5,700 | 17.31 | 31.38 | +| Verifier runtime (seconds) | 5,699 | 5.87 | 6.22 | +| Whole-attempt runtime (seconds) | 5,700 | 31.71 | 47.36 | -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-Coder-30B-A3B-Instruct | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"repetition_penalty": 1.05, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +## Plots + +![Trace measurements](plots/all-models-trace-distributions.png) -The table reports settings found in **1 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](../metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +![Trace outcomes](plots/all-models-trace-outcomes.png) + +![Most frequent commands](plots/all-models-trace-commands.png) diff --git a/nl2bash_v3/README.md b/nl2bash_v3/README.md index ce0e922e5fdafab8ad7d5af1943ed89ae98c522f..e6a6cb1175fe138fe5d7e983ed219a03274ac6bf 100644 --- a/nl2bash_v3/README.md +++ b/nl2bash_v3/README.md @@ -2,101 +2,81 @@ ## Dataset stats -- **Original tasks:** 1,497 -- **Kept in v3:** 104 newly solved tasks -- **Available separately in v2:** 938 -- **Combined v2 + v3:** 1,042 unique tasks -- **Newly solved:** 104 of 487 previously unsolved tasks (21.36%) -- **Changed task payloads:** 0 -- **Archived:** 455 (383 still unsolved and 72 earlier exclusions) -- **Pending attempts:** 0 -- **New evaluation:** 487 tasks, 974 attempts, 974 saved conversations -- **Historical conversations:** available separately in [nl2bash_v2](../nl2bash_v2/README.md) -- **Images:** 1 before, 1 now +- **Tasks:** 487 had no successful attempt in v2; all 487 evaluated for v3; 104 newly kept; 383 still without success. +- **Teacher evaluation:** Qwen/Qwen3.8-27B (thinking off), 2 attempts per task (pass@2). +- **Coverage:** 487 tasks evaluated; 974 evaluation attempts. +- **Saved conversations:** 974 of 974 published attempt records, including successful and unsuccessful attempts. -The 938 [v2 tasks](../nl2bash_v2/README.md) remain available separately. V3 adds tasks solved at least once by Qwen/Qwen3.8-27B in two attempts. The v1 source revision is `3abfc63260ef8ab59ac61274450310f650585bc9`. - -## Dataset characteristics - -NL2Bash tasks translate natural-language requests into shell commands for file operations and system administration. Agents work with seeded files in `/workspace` and capture command output in `/output/command_capture.txt`. Intended transfer includes Terminal-Bench 2.0, OpenThoughts-TBLite 2.0 and Harbor-Index. - -Task payloads and pinned image references are unchanged. The earlier Git-history and timestamp leak-check limitation remains unresolved. Observed success is a solvability filter, not proof of verifier correctness: inspected zero-score examples included output-format mismatches and faulty expected output. These new results do not certify completion of all datasource checks. The run recorded host networking after isolated network namespace creation failed. - -## New evaluation - -Qwen/Qwen3.8-27B with Terminus 2 evaluated exactly the 487 v2 tasks with no successful Qwen/Qwen3-Coder-30B-A3B-Instruct attempt. The four-task pilot and 483-task full run were disjoint. All 22 stage-6 reports and 22 stage-7 reports are complete. Each task received two terminal attempts. Jobs requested one H200, 32 CPUs including four serving CPUs, and four concurrent trials. - -Agent and verifier timeouts are 600 seconds each in all 974 saved trial configurations; task limits were unchanged. [Contracts](metadata/full-contract.json), [settings](metadata/evaluation-settings.json), and the [saved trial configuration](metadata/sample-trial-config.json) preserve the evidence. +We keep tasks with at least one verifier-successful attempt from the new model. `tasks.parquet` contains only newly kept tasks; tasks already kept in v2 remain separate. ## Evaluation settings -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | Off | 2 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | presence_penalty=1.5; repetition_penalty=1 | -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3.8-27B | 32768 | 8192 | present in saved request kwargs; HTTP payload not audited | off | {"presence_penalty": 1.5, "repetition_penalty": 1.0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -Configured limits are distinct from measured token usage. +**Time limits:** agent 600 seconds; verifier 600 seconds (all 974 saved configurations; limits vary by task where multiple values are shown). -The 8,192-token response cap is present in saved `llm_call_kwargs.max_tokens`; HTTP requests were not independently audited. Historical v2 settings and measurements remain in the v2 README. +**Output-limit evidence:** All 974 saved configurations contain an explicit 8,192-token request cap. HTTP payloads and server enforcement were not audited. -## New success distribution +Network isolation was requested, but the run fell back to host networking when network namespace creation failed; agents could have reached the network. -These rates cover the 487 previously unsolved tasks, not the 104 success-selected tasks published in v3. +## Results -| Metric | Rate | -| --- | ---: | -| pass@1 | 16.12% | -| pass@2 | 21.36% | +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. Only the new v3 evaluation is included. -| Successful attempts out of two | Tasks | -| ---: | ---: | -| 0 | 383 | -| 1 | 51 | -| 2 | 53 | +Observed success is a solvability filter, not proof of verifier correctness. Inspected zero-score examples included output-format mismatches and faulty expected output. -![New pass rates](plots/pass-rates.png) +| Successful attempts | Qwen/Qwen3.8-27B (487 tasks) | +| --- | ---: | +| 0/2 | 383 | +| 1/2 | 51 | +| 2/2 | 53 | -![New success distribution](plots/solved-counts.png) +| Model | Evaluated tasks | Pass@1 | Pass@2 | +| --- | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | 487 | 16.12% | 21.36% | -## Trajectories and runtime +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. -All 974 new conversations are saved, with 157 successful attempts and 973 numeric verifier rewards. One terminal attempt ended with a TmuxBatchProtocolError and has a null raw verifier reward and counts as zero under the recorded reward policy. No missing attempts were fabricated and no conversations were removed. Stage-7 completion is not a claim that every saved conversation is suitable for training. +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -The 5,700 historical conversations remain in v2. This folder contains the 974 new attempts; `retained_v3` indicates membership in the incremental task set. +## Files and usage -| New-run measurement | Median | -| --- | ---: | -| Agent turns | 5 | -| Input tokens per attempt | 9,957 | -| Output tokens per attempt | 1,175 | -| Agent runtime, seconds | 28 | -| Verifier runtime, seconds | 6 | +`tasks.parquet` contains 104 newly kept tasks only. Tasks already kept in v2 are not duplicated. -Medians use recorded measurements. Token totals sum calls within each attempt. +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. -[Trajectory statistics](trace_stats/README.md) contain new-evaluation-only token, turn, runtime and outcome measurements, plots and per-attempt data. Historical statistics are linked through v2. +Success-count columns: `coder_successes`, `qwen38_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. -## Files +`archive.parquet` preserves excluded task payloads and reasons, including earlier exclusions where applicable. -- [tasks.parquet](tasks.parquet): 104 newly solved tasks, preserving existing columns and adding Qwen/Qwen3.8-27B attempt and success counts. -- [archive.parquet](archive.parquet): 455 excluded tasks, preserving task payloads and prior metadata. -- [Selection](metadata/selection.json): incremental membership, combined membership and per-task evidence. -- [Provenance](metadata/provenance.json): source revisions and report/archive hashes. -- [Trace manifest](traces/manifest.json): all new conversations with file hashes. +```python +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -## Incremental release +path = hf_hub_download( + "FWeindel/validated-tasks", + "nl2bash_v3/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("qwen38_successes", "=", 1)]) +``` -V3 contains only tasks newly solved after v2. It has no task-ID overlap with [v2](../nl2bash_v2/README.md). This is a harder subset relative to the earlier evaluation, not a claim of intrinsic difficulty. New-evaluation traces include both solved and unsolved tasks so the reported pass rates remain reproducible. +Combine the 938 tasks in `nl2bash_v2/tasks.parquet` with these 104 newly kept tasks for 1,042 unique tasks. Allow for different columns and retain each source file's image manifest when staging images. ```python -from huggingface_hub import hf_hub_download import pyarrow as pa -import pyarrow.parquet as pq def load(version): - path = hf_hub_download("FWeindel/validated-tasks", - f"nl2bash_{version}/tasks.parquet", repo_type="dataset") + path = hf_hub_download( + "FWeindel/validated-tasks", + f"nl2bash_{version}/tasks.parquet", + repo_type="dataset", + ) return pq.read_table(path) v2, v3 = load("v2"), load("v3") @@ -104,4 +84,14 @@ assert set(v2["path"].to_pylist()).isdisjoint(v3["path"].to_pylist()) combined = pa.concat_tables([v2, v3], promote_options="permissive") ``` -The example combines task rows; each source file retains its own image manifest. Use those source manifests when staging images. [Layout and provenance](metadata/layout.json) records the pinned baseline and task counts. Earlier cumulative v3 releases remain accessible through their HF commit history. +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. + +## Trace statistics + +Of 974 recorded attempts, 974 have saved conversations and 0 do not. 1 attempt has no raw verifier reward. Missing measurements remain missing. + +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. + +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. Historical attempts remain in v2. + +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/nl2bash_v3/plots/readme-pass-at-k.png b/nl2bash_v3/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..eb987da2adb2c3c82998ad28828aee05ad55ad1f --- /dev/null +++ b/nl2bash_v3/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8763a25a4ea0a4ebfd77f124da3f92734482b48b92586a19b9abfffff987dd17 +size 35955 diff --git a/nl2bash_v3/trace_stats/README.md b/nl2bash_v3/trace_stats/README.md index 40ac898375b0cf532c0dd77bb93a76113dc0c781..f1c3decdb0aaeed47ec5997bdca60fcef1899295 100644 --- a/nl2bash_v3/trace_stats/README.md +++ b/nl2bash_v3/trace_stats/README.md @@ -1,52 +1,53 @@ -# Trajectory statistics +# NL2Bash v3 trace statistics -974 attempts across 487 tasks of NL2Bash v3, Qwen/Qwen3.8-27B evaluation, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3.8-27B | -| --- | ---: | -| Recorded attempts | 974 | -| Median agent turns | 5 | -| Median input tokens per attempt | 9,957 | -| Median output tokens per attempt | 1,175 | -| Median agent runtime (seconds) | 28 | -| Median verifier runtime (seconds) | 6 | +Of 974 recorded attempts, 974 have saved conversations and 0 do not. 1 attempt has no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. Historical attempts remain in v2. -![Trace measurements](plots/all-models-trace-distributions.png) +Token totals sum calls across an attempt and are not the context window. -![Trace outcomes](plots/all-models-trace-outcomes.png) +Measurements cover 487 tasks, including kept and unselected tasks, with each attempt counted once. -![Most frequent commands](plots/all-models-trace-commands.png) +## Evaluation settings -Commands are counted as submitted, not as successful executions. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | Off | 2 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | presence_penalty=1.5; repetition_penalty=1 | -Per-model plots: [Qwen/Qwen3.8-27B commands](plots/qwen3.8-27b-trace-commands.png), [Qwen/Qwen3.8-27B distributions](plots/qwen3.8-27b-trace-distributions.png), [Qwen/Qwen3.8-27B outcomes](plots/qwen3.8-27b-trace-outcomes.png). +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Data +**Time limits:** agent 600 seconds; verifier 600 seconds (all 974 saved configurations; limits vary by task where multiple values are shown). -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +**Output-limit evidence:** All 974 saved configurations contain an explicit 8,192-token request cap. HTTP payloads and server enforcement were not audited. -Attempts without a saved conversation: 0; without a verifier reward: 1. Their missing measurements stay empty. +Network isolation was requested, but the run fell back to host networking when network namespace creation failed; agents could have reached the network. +## Measurements -Evaluation budgets: see the [dataset evaluation setup](../README.md). Per-call context and output-limit enforcement are not established by token-usage statistics. +### Qwen/Qwen3.8-27B +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 974 | 5 | 9 | +| Input tokens per attempt | 974 | 9,957 | 28,272 | +| Output tokens per attempt | 974 | 1,175 | 2,242 | +| Largest recorded step, prompt plus output tokens | 974 | 2,976 | 5,619 | +| Agent runtime (seconds) | 974 | 27.79 | 59.3 | +| Verifier runtime (seconds) | 973 | 6.13 | 6.87 | +| Whole-attempt runtime (seconds) | 974 | 42.65 | 73.33 | -## Evaluation settings +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3.8-27B | 32768 | 8192 | present in saved request kwargs; HTTP payload not audited | off | {"presence_penalty": 1.5, "repetition_penalty": 1.0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +## Plots -Configured limits are distinct from measured token usage. +![Trace measurements](plots/all-models-trace-distributions.png) -Only the 974 new v3 attempts are summarized here. See [v2](../../nl2bash_v2/README.md) for historical statistics. Agent and verifier limits are 600 seconds in all 974 saved trial configurations. [Settings evidence](../metadata/evaluation-settings.json). +![Trace outcomes](plots/all-models-trace-outcomes.png) -This folder reports the new evaluation only. Task membership is incremental: tasks already in v2 are not duplicated in v3. Historical conversations remain in v2. See the [release overview](../README.md) for counts and combination instructions. +![Most frequent commands](plots/all-models-trace-commands.png) diff --git a/scaleswe_v2/README.md b/scaleswe_v2/README.md index 29e902d0527e6fa64fb2b2d323144e8723272c51..b8cd28e069a2559151c7f8b893f625fc43411db2 100644 --- a/scaleswe_v2/README.md +++ b/scaleswe_v2/README.md @@ -1,89 +1,38 @@ # Scale-SWE v2 -Results snapshot: **2026-10-26T15:55:39.135106+00:00**. - ## Dataset stats -- **v1 tasks:** 16,786 -- **Kept:** 2,655 (solved at least once by either model, including historical recovery evidence) -- **Not selected:** 14,131 -- **Incomplete Qwen/Qwen3-30B-A3B-Instruct-2507 tasks:** 0 -- **Recorded attempts:** 162,800 -- **Available trajectories:** 162,800 -- **Task payload changes:** 0 -- **v1 base revision:** `2aaa0204c66770f87f803485b337a6e363edf647` -- **Newly kept in this refresh:** 0 - -Scale-SWE contains software engineering tasks from Python repositories. Tasks preserve the v1 payloads and are selected when either teacher has solved them at least once. - -## Evaluation settings - -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. - -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | -| Qwen/Qwen3-30B-A3B | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | on | {"min_p": 0, "temperature": 0.6, "top_k": 20, "top_p": 0.95} | - -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. - -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. - -The table reports settings found in **8 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. - -### Completed retry settings - -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. - -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 (36-task retry) | 32768 | 8192 | top-level only; explicit request cap not forwarded by Terminus | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | - -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. - -All 288 saved retry configurations were inspected. Explicit output caps remain top-level only in these Terminus configurations; the actual fallback cap is unknown. Resolved limits: 1,200 seconds for the agent, 1,800 seconds for the verifier, 600 seconds for environment build, 360 seconds for agent setup, and 30 seconds for cleanup. No whole-attempt timeout is recorded. [Retry settings evidence](metadata/retry-settings-20261021.json). - -### Final retry settings +- **Tasks:** 16,786 in v1; 2,655 kept in v2. +- **Teacher evaluation:** Qwen/Qwen3-30B-A3B-Instruct-2507 (thinking off), 8 attempts per task (pass@8); Qwen/Qwen3-30B-A3B (thinking on), 8 attempts per task (pass@8). +- **Coverage:** 16,786/16,786 tasks evaluated; 162,800 evaluation attempts. +- **Saved conversations:** 162,800 of 162,800 published attempt records, including successful and unsuccessful attempts. -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +We keep tasks with at least one verifier-successful attempt from any evaluated model. -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | top-level only; explicit request cap not forwarded by Terminus | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +The published set also preserves 6 tasks from the earlier selection that have no success in the current eight-attempt evidence. -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. - -All 96 saved configurations from the final 12-task retry were inspected. The intended output cap is top-level only and was not forwarded by the affected Terminus path; the actual fallback cap is unknown. Resolved limits: 1,200 seconds for the agent, 1,800 seconds for the verifier, 600 seconds for environment build, 360 seconds for agent setup, and 30 seconds for cleanup. No whole-attempt timeout is recorded. [Settings evidence](metadata/retry-settings-20261026.json) and [saved reports and configurations](metadata/retry-reports-20261026.tar.gz). - -## Pass@8 collection status - -| Model | Tasks attempted | Recorded attempts | Tasks with all 8 outcomes | Tasks solved | -| --- | ---: | ---: | ---: | ---: | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 16,786 | 134,288 | 16,786 | 2,515 | -| Qwen/Qwen3-30B-A3B (thinking) | 3,564 | 28,512 | 3,564 | 375 | - -The teacher is `Qwen/Qwen3-30B-A3B-Instruct-2507`, using Terminus-2, eight attempts per task, temperature 0.7, top-p 0.8, top-k 20, a 32,768-token context and an intended 8,192-token response limit (see evaluation settings). The existing Qwen/Qwen3-30B-A3B (thinking) collection remains unchanged. +## Evaluation settings -Recorded terminal errors receive reward **0**, following the collection policy `terminal-errors-zero-v2`. Raw verifier rewards and error details remain in the trace evidence. Unfinished or unrun attempts remain missing. Racing copies are counted once. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 8 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | +| Qwen/Qwen3-30B-A3B | Terminus-2 | On | 8 | 0.6 | 0.95 | 20 | 32,768 | 8,192 | min_p=0 | -All 16,786 tasks have eight recorded outcomes from `Qwen/Qwen3-30B-A3B-Instruct-2507`. The final 12 tasks produced 96 verifier-graded outcomes, all zero, so the kept set is unchanged. [Final retry evidence](metadata/retry-results-20261026.json). `Qwen/Qwen3-30B-A3B` (thinking) remains a stopped, partial collection. +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Pass@k on complete task cohorts +**Time limits:** agent 1,200 seconds; verifier 1,800 seconds (sampled configurations; limits vary by task where multiple values are shown). -Qwen/Qwen3-30B-A3B-Instruct-2507 covers all 16,786 v1 tasks with eight outcomes each. Qwen/Qwen3-30B-A3B (thinking) covers only its stopped partial collection of 3,564 tasks. Terminal errors are scored zero under the recorded policy. +**Output-limit evidence:** The 8 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -| Model | Tasks | pass@1 | pass@2 | pass@4 | pass@8 | -| --- | ---: | ---: | ---: | ---: | ---: | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 16,786 | 4.10% | 6.81% | 10.47% | 14.98% | -| Qwen/Qwen3-30B-A3B (thinking) | 3,564 | 2.64% | 4.49% | 7.13% | 10.52% | +## Results -![Pass rates and successful-attempt counts](trace_stats/plots/pass-at-k.png) +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. -### Successful attempts among kept tasks +Qwen/Qwen3-30B-A3B-Instruct-2507 covers all 16,786 v1 tasks. Qwen/Qwen3-30B-A3B covers only 3,564 tasks from its stopped collection; 13,222 v1 tasks were not evaluated by that model and are not counted as 0/8. The current eight-attempt evidence solves 2,649 unique tasks; 6 additional tasks remain kept from the earlier selection without a success in this evidence. -| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 | Qwen/Qwen3-30B-A3B (thinking) | +| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 (16,786 tasks) | Qwen/Qwen3-30B-A3B (3,564 tasks) | | --- | ---: | ---: | -| 0/8 | 140 | 2,280 | +| 0/8 | 14,271 | 3,189 | | 1/8 | 1,245 | 202 | | 2/8 | 512 | 78 | | 3/8 | 307 | 40 | @@ -93,28 +42,44 @@ Qwen/Qwen3-30B-A3B-Instruct-2507 covers all 16,786 v1 tasks with eight outcomes | 7/8 | 45 | 2 | | 8/8 | 26 | 5 | -Counts of zero successful attempts can include tasks without completed attempts for that model. Coverage is shown above. +| Model | Evaluated tasks | Pass@1 | Pass@2 | Pass@4 | Pass@8 | +| --- | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | 16,786 | 4.10% | 6.81% | 10.47% | 14.98% | +| Qwen/Qwen3-30B-A3B | 3,564 | 2.64% | 4.49% | 7.13% | 10.52% | + +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. + +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) + +## Files and usage -## Reproduce and select tasks +`tasks.parquet` contains 2,655 kept tasks. -`patch.py` rebuilds `tasks.parquet` from v1 at revision `2aaa0204c66770f87f803485b337a6e363edf647` and `metadata/selection.json`, verifying the source SHA-256. Task payloads are preserved; trace hashes identify the version evaluated. +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. + +Success-count columns: `instruct_successes`, `thinking_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. ```python +from huggingface_hub import hf_hub_download import pyarrow.parquet as pq -subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)]) -``` -The `*_graded_attempts` columns count scored outcomes under the collection policy; `*_verifier_graded_attempts` count actual verifier grades. `*_missing_attempts` counts missing scored outcomes out of eight. +path = hf_hub_download( + "FWeindel/validated-tasks", + "scaleswe_v2/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("instruct_successes", "=", 1)]) +``` -## Evidence and trajectories +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. -The combined `traces/` files contain **162,800 unique attempts**, including kept and excluded tasks. Each row stores the outcome reward, raw `verifier_reward`, `reward_origin`, original result and available trajectory. `retained_v2` matches this selection. +## Trace statistics -[Final retry evidence](metadata/retry-results-20261026.json) records the last 12 task evaluations; [previous retry evidence](metadata/retry-results-20261021.json) covers the earlier 36; [earlier selection evidence](metadata/new-evidence-20261020.json) is preserved. [Trace statistics](trace_stats/README.md) include updated per-model outcomes, token usage, tool usage, runtime and plots. All trace files and statistics describe this snapshot. +Of 162,800 recorded attempts, 162,800 have saved conversations and 0 do not. 1,061 attempts have no raw verifier reward. Missing measurements remain missing. -## Runtime +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -- **Versions:** Apptainer **1.5.4** on Helma and **1.5.2** on Capella, Python **3.12.14**, Harbor **0.8.1**. -- **Implementation:** [Marin Harbor `7faf878c14b6`](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) + [our runtime patches](https://github.com/MLI-lab/OT-next-data/tree/main/harbor_patches). +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -Not yet checked with the repository-leak checks added on 2026-10-17 (leftover git history and file timestamps that could reveal the solution). +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/scaleswe_v2/plots/readme-pass-at-k.png b/scaleswe_v2/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..aa30872aa2f525636f828f220b9a6d59c5e605f0 --- /dev/null +++ b/scaleswe_v2/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d00d8cd5e74b88ede7bbd6bc483e7faf3b000f959412c47ba6ec2d604ca413ad +size 50764 diff --git a/scaleswe_v2/trace_stats/README.md b/scaleswe_v2/trace_stats/README.md index 1a3ba3be303dcfe98b41a83ab2a46a0ba5a98a46..9951751b8225a0da24467efaaa30d969ec66eec2 100644 --- a/scaleswe_v2/trace_stats/README.md +++ b/scaleswe_v2/trace_stats/README.md @@ -1,78 +1,64 @@ -# Trajectory statistics +# Scale-SWE v2 trace statistics -162,800 attempts across 16,786 tasks of Scale SWE v2, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3-30B-A3B-Instruct-2507 | Qwen/Qwen3-30B-A3B (thinking) | -| --- | ---: | ---: | -| Recorded attempts | 134,288 | 28,512 | -| Median agent turns | 15 | 17 | -| Median input tokens per attempt | 94,487 | 128,586 | -| Median output tokens per attempt | 5,203 | 14,999 | -| Median agent runtime (seconds) | 69 | 139 | -| Median verifier runtime (seconds) | 8 | 8 | +Of 162,800 recorded attempts, 162,800 have saved conversations and 0 do not. 1,061 attempts have no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots - -![Trace measurements](plots/all-models-trace-distributions.png) - -![Trace outcomes](plots/all-models-trace-outcomes.png) - -![Most frequent commands](plots/all-models-trace-commands.png) - -Commands are counted as submitted, not as successful executions. - -Per-model plots: [Qwen/Qwen3-30B-A3B-Instruct-2507 commands](plots/qwen3-30b-instruct-2507-trace-commands.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 distributions](plots/qwen3-30b-instruct-2507-trace-distributions.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 outcomes](plots/qwen3-30b-instruct-2507-trace-outcomes.png), [Qwen/Qwen3-30B-A3B (thinking) commands](plots/qwen3-30b-thinking-trace-commands.png), [Qwen/Qwen3-30B-A3B (thinking) distributions](plots/qwen3-30b-thinking-trace-distributions.png), [Qwen/Qwen3-30B-A3B (thinking) outcomes](plots/qwen3-30b-thinking-trace-outcomes.png). +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -## Data +Token totals sum calls across an attempt and are not the context window. -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. - -Attempts without a saved conversation: 0; without a verifier reward: 1,061. Their missing measurements stay empty. - - -Evaluation budgets: see the [dataset evaluation setup](../README.md). Per-call context and output-limit enforcement are not established by token-usage statistics. +Measurements cover 16,786 tasks, including kept and unselected tasks, with each attempt counted once. ## Evaluation settings -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. - -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | -| Qwen/Qwen3-30B-A3B | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | on | {"min_p": 0, "temperature": 0.6, "top_k": 20, "top_p": 0.95} | - -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 8 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | +| Qwen/Qwen3-30B-A3B | Terminus-2 | On | 8 | 0.6 | 0.95 | 20 | 32,768 | 8,192 | min_p=0 | -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -The table reports settings found in **8 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](../metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +**Time limits:** agent 1,200 seconds; verifier 1,800 seconds (sampled configurations; limits vary by task where multiple values are shown). -### Completed retry settings +**Output-limit evidence:** The 8 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +## Measurements -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 (36-task retry) | 32768 | 8192 | top-level only; explicit request cap not forwarded by Terminus | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +### Qwen/Qwen3-30B-A3B-Instruct-2507 -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 134,288 | 15 | 85 | +| Input tokens per attempt | 134,288 | 94,487 | 1,345,615 | +| Output tokens per attempt | 134,288 | 5,203 | 66,309 | +| Largest recorded step, prompt plus output tokens | 134,280 | 12,223 | 26,189 | +| Agent runtime (seconds) | 134,288 | 69.37 | 930.85 | +| Verifier runtime (seconds) | 134,096 | 7.69 | 19.52 | +| Whole-attempt runtime (seconds) | 134,288 | 98.15 | 960.61 | -All 288 saved retry configurations were inspected. Explicit output caps remain top-level only in these Terminus configurations; the actual fallback cap is unknown. Resolved limits: 1,200 seconds for the agent, 1,800 seconds for the verifier, 600 seconds for environment build, 360 seconds for agent setup, and 30 seconds for cleanup. No whole-attempt timeout is recorded. [Retry settings evidence](../metadata/retry-settings-20261021.json). +### Qwen/Qwen3-30B-A3B +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 28,512 | 17 | 112 | +| Input tokens per attempt | 28,512 | 128,586 | 1,713,359 | +| Output tokens per attempt | 28,512 | 14,999 | 125,878 | +| Largest recorded step, prompt plus output tokens | 28,510 | 14,097 | 25,763 | +| Agent runtime (seconds) | 28,512 | 138.74 | 1,200.01 | +| Verifier runtime (seconds) | 28,433 | 7.64 | 22.49 | +| Whole-attempt runtime (seconds) | 28,512 | 168.29 | 1,229.29 | -### Final retry settings +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | top-level only; explicit request cap not forwarded by Terminus | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +## Plots -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +![Trace measurements](plots/all-models-trace-distributions.png) -All 96 saved configurations from the final 12-task retry were inspected. The intended output cap is top-level only and was not forwarded by the affected Terminus path; the actual fallback cap is unknown. Resolved limits: 1,200 seconds for the agent, 1,800 seconds for the verifier, 600 seconds for environment build, 360 seconds for agent setup, and 30 seconds for cleanup. No whole-attempt timeout is recorded. [Settings evidence](../metadata/retry-settings-20261026.json) and [saved reports and configurations](../metadata/retry-reports-20261026.tar.gz). +![Trace outcomes](plots/all-models-trace-outcomes.png) +![Most frequent commands](plots/all-models-trace-commands.png) diff --git a/seta_v2/README.md b/seta_v2/README.md index a0cac90ec86ce8e69e430a1f4149fe5e47a2146b..aa26c4b48527812ae81c4ce545b10b12ab5f0f3d 100644 --- a/seta_v2/README.md +++ b/seta_v2/README.md @@ -2,140 +2,94 @@ ## Dataset stats -- **v1 tasks:** 3,107 -- **Kept:** 2,246, solved at least once by either model, including historical recovery results. -- **Excluded:** 861, no recorded success. -- **Evaluated:** 3,107 tasks with 16 completed outcomes per model; 99,424 evaluation attempts. -- **Pending evaluation attempts:** 0. -- **Qwen/Qwen3-30B-A3B-Instruct-2507 (thinking off):** pass@16 **69.81%**, solving 2,169 of 3,107 tasks. -- **Qwen/Qwen3-30B-A3B (thinking on):** pass@16 **47.67%**, solving 1,481 of 3,107 tasks. -- **Combined trace records:** 112,008, including historical attempts. -- **Saved conversations:** 111,700. +- **Tasks:** 3,107 in v1; 2,246 kept in v2. +- **Teacher evaluation:** Qwen/Qwen3-30B-A3B-Instruct-2507 (thinking off), 16 attempts per task (pass@16); Qwen/Qwen3-30B-A3B (thinking on), 16 attempts per task (pass@16). +- **Coverage:** 3,107/3,107 tasks evaluated; 99,424 evaluation attempts. +- **Saved conversations:** 111,700 of 112,008 published attempt records, including successful and unsuccessful attempts. -## Evaluation settings - -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. - -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | on | {"min_p": 0, "temperature": 0.6, "top_k": 20, "top_p": 0.95} | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | - -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. - -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +We keep tasks with at least one verifier-successful attempt from any evaluated model. -The table reports settings found in **7 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +The kept set includes historical successes; the evaluation and historical trace counts are distinguished below. -## Dataset characteristics and changes from v1 - -SETA covers terminal workflows for data science, debugging, system administration and data processing. v2 selects tasks with at least one recorded success and keeps their pinned v1 payloads unchanged. 2,233 tasks were solved in the completed 16-attempt collection; another 13 are kept from historical recovery successes. - -## Evaluation setup and coverage - -Both models used Terminus 2. Saved gap-fill contracts specify a 32,768-token context and an intended 8,192-token output limit per model call (see evaluation settings). Qwen/Qwen3-30B-A3B-Instruct-2507 sampling uses temperature 0.7 and top-p 0.8; Qwen/Qwen3-30B-A3B (thinking) uses temperature 0.6 and top-p 0.95. Both use top-k 20 and min-p 0. Task-specific agent limits range from 180 to 3,600 seconds and verifier limits from 120 to 1,800 seconds. [Saved setup](metadata/evaluation-setup.json) and [recorded limits](metadata/runtime-resources.json). - -| Model | Tasks attempted | Recorded outcomes | Raw verifier scores | Tasks solved | -| --- | ---: | ---: | ---: | ---: | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 3,107 | 49,712 | 49,472 | 2,169 | -| Qwen/Qwen3-30B-A3B (thinking) | 3,107 | 49,712 | 49,459 | 1,481 | +## Evaluation settings -Both collections are complete. Under the existing `terminal-errors-zero-v2` policy, the 240 Qwen/Qwen3-30B-A3B-Instruct-2507 and 253 Qwen/Qwen3-30B-A3B (thinking) attempts that ended with errors count as unsuccessful outcomes for pass@k. Their raw verifier rewards remain `null`, and their exceptions are preserved. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B | Terminus-2 | On | 16 | 0.6 | 0.95 | 20 | 32,768 | 8,192 | min_p=0 | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 16 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | -Historical recovery attempts are included in the combined traces but are not added to the 16-attempt evaluation. Qwen/Qwen3.5-27B retry results are separate and are not included. +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Pass@k and success distributions +**Time limits:** agent 3,600 seconds; verifier 120, 900 seconds (sampled configurations; limits vary by task where multiple values are shown). The broader saved setup reports task-specific agent limits of 180 to 3,600 seconds and verifier limits of 120 to 1,800 seconds. -| Model | pass@1 | pass@8 | pass@16 | -| --- | ---: | ---: | ---: | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 35.94% | 63.38% | 69.81% | -| Qwen/Qwen3-30B-A3B (thinking) | 19.99% | 41.06% | 47.67% | +**Output-limit evidence:** The 7 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -Each metric covers all 3,107 tasks. pass@8 uses the standard estimator from 16 attempts; pass@16 is the fraction solved at least once. +## Results -![Successful-attempt distributions](metadata/plots/success-distributions.png) +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. -
Successful attempts among the 2,246 kept tasks +The pass@16 evaluation solved 2,233 unique tasks. Another 13 tasks are kept because of historical successes, giving 2,246 kept tasks. The 12,584 historical attempts are preserved in traces but excluded from these fixed-budget results. -| Successes out of 16 | Qwen/Qwen3-30B-A3B-Instruct-2507 | Qwen/Qwen3-30B-A3B (thinking) | +| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 (3,107 tasks) | Qwen/Qwen3-30B-A3B (3,107 tasks) | | --- | ---: | ---: | -| 0 | 77 | 765 | -| 1 | 266 | 301 | -| 2 | 202 | 169 | -| 3 | 126 | 104 | -| 4 | 133 | 95 | -| 5 | 127 | 77 | -| 6 | 95 | 64 | -| 7 | 103 | 73 | -| 8 | 86 | 57 | -| 9 | 82 | 57 | -| 10 | 95 | 66 | -| 11 | 114 | 66 | -| 12 | 91 | 66 | -| 13 | 123 | 67 | -| 14 | 111 | 60 | -| 15 | 155 | 76 | -| 16 | 260 | 83 | - -
- -
Results by task family (all evaluated tasks) - -| Task family | Tasks | Qwen/Qwen3-30B-A3B-Instruct-2507 pass@1 | Qwen/Qwen3-30B-A3B-Instruct-2507 pass@8 | Qwen/Qwen3-30B-A3B-Instruct-2507 pass@16 | Qwen/Qwen3-30B-A3B (thinking) pass@1 | Qwen/Qwen3-30B-A3B (thinking) pass@8 | Qwen/Qwen3-30B-A3B (thinking) pass@16 | -| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | -| ask_ubuntu | 1,132 | 54.30% | 80.46% | 85.42% | 37.38% | 66.19% | 72.88% | -| kaggle_notebook | 645 | 16.52% | 47.03% | 55.81% | 1.14% | 6.31% | 10.08% | -| nl2bash | 214 | 10.40% | 30.39% | 37.38% | 2.75% | 10.69% | 15.42% | -| stack_overflow | 766 | 36.94% | 64.62% | 70.63% | 17.64% | 43.10% | 51.44% | -| unix_linux_se | 350 | 25.77% | 55.69% | 63.14% | 14.14% | 37.91% | 46.86% | - -
- -| Observed reward pattern (all tasks) | Qwen/Qwen3-30B-A3B-Instruct-2507 | Qwen/Qwen3-30B-A3B (thinking) | -| --- | ---: | ---: | -| Always solved | 260 | 83 | -| Always unsuccessful | 938 | 1,626 | -| Varying | 1,909 | 1,398 | -| Constant partial credit | 0 | 0 | +| 0/16 | 938 | 1,626 | +| 1/16 | 266 | 301 | +| 2/16 | 202 | 169 | +| 3/16 | 126 | 104 | +| 4/16 | 133 | 95 | +| 5/16 | 127 | 77 | +| 6/16 | 95 | 64 | +| 7/16 | 103 | 73 | +| 8/16 | 86 | 57 | +| 9/16 | 82 | 57 | +| 10/16 | 95 | 66 | +| 11/16 | 114 | 66 | +| 12/16 | 91 | 66 | +| 13/16 | 123 | 67 | +| 14/16 | 111 | 60 | +| 15/16 | 155 | 76 | +| 16/16 | 260 | 83 | -These patterns describe the observed samples under the stated error policy. +| Model | Evaluated tasks | Pass@1 | Pass@2 | Pass@4 | Pass@8 | Pass@16 | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | 3,107 | 35.94% | 46.31% | 55.54% | 63.38% | 69.81% | +| Qwen/Qwen3-30B-A3B | 3,107 | 19.99% | 27.23% | 34.28% | 41.06% | 47.67% | -## Traces and trajectory statistics +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. -The combined traces contain **112,008 unique attempts**, including 12,584 historical attempts outside the evaluation. **308** have no saved conversation and **5,789** have no raw verifier reward. All attempt records are preserved, including errors; a saved conversation is not a claim of a successful or quality-filtered training trajectory. +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -Every row includes the original result JSON, available trajectory, outcome reward, raw verifier reward and current `retained_v2` flag. `in_pass16_collection` identifies the attempts used in the evaluation and task result columns. +## Files and usage -215 kept tasks have historical attempts on an earlier task version. The task Parquet contains pinned v1 payloads; traces preserve evaluated content hashes. +`tasks.parquet` contains 2,246 kept tasks. -[Recomputed trace statistics](trace_stats/README.md) cover all combined attempts, including outcomes, commands, tokens and runtime. [Result metadata](metadata/results-20261020.json) records the completed evaluation and added-report provenance. +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. -## Runtime and tokens +Success-count columns: `instruct_successes`, `thinking_successes`, `recovery_instruct_successes`, `recovery_thinking_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. -| Median per attempt (all combined traces) | Qwen/Qwen3-30B-A3B-Instruct-2507 | Qwen/Qwen3-30B-A3B (thinking) | -| --- | ---: | ---: | -| Agent turns | 12.0 | 18.0 | -| Input tokens | 76,716.0 | 150,250.0 | -| Output tokens | 6,343.0 | 25,012.0 | -| Peak context tokens | 13,144.0 | 17,184.0 | -| Agent seconds | 145.5 | 477.3 | -| Verifier seconds | 4.7 | 4.6 | -| Trial seconds | 185.0 | 518.3 | +```python +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -Medians use recorded measurements only; sample counts and percentiles are in the linked statistics. +path = hf_hub_download( + "FWeindel/validated-tasks", + "seta_v2/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("instruct_successes", "=", 1)]) +``` -Measured agent-environment peak memory has a median of **1,401.8 MB** across **6,024 Qwen/Qwen3-30B-A3B-Instruct-2507 attempts** and **1,356.9 MB** across **6,246 Qwen/Qwen3-30B-A3B (thinking) attempts**. [Resource distributions](metadata/runtime-resources.json) and [per-task peaks](metadata/per-task-memory.parquet) cover attempts with saved measurements only. +215 kept tasks have historical attempts on an earlier task version; the published file contains pinned v1 payloads. -- **Versions:** Apptainer **1.5.4**, Python **3.12.14**, Harbor **0.8.1**. -- **Implementation:** [Marin Harbor](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) + [our runtime patches](https://github.com/MLI-lab/OT-next-data/tree/main/harbor_patches). +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. -Not yet checked with the repository-leak checks added on 2026-10-17 (leftover git history and file timestamps that could reveal the solution). +## Trace statistics -## Reproduce and select tasks +Of 112,008 recorded attempts, 111,700 have saved conversations and 308 do not. 5,789 attempts have no raw verifier reward. Missing measurements remain missing. These counts cover all combined traces, including 12,584 historical attempts outside the pass@16 evaluation. -`patch.py` rebuilds `tasks.parquet` from the pinned v1 source recorded in [selection and provenance](metadata/selection.json), verifying the source SHA-256. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -```python -import pyarrow.parquet as pq -subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)]) -``` +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. + +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/seta_v2/plots/readme-pass-at-k.png b/seta_v2/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..bc0ef48645e4ed406aef6583f2d15de2ed426b57 --- /dev/null +++ b/seta_v2/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56b2f8704ff4639ac134b7985ffefef94e2f121098ec8d13c26f541d360e8343 +size 57438 diff --git a/seta_v2/trace_stats/README.md b/seta_v2/trace_stats/README.md index c03e985abe751cae50853175d77f49f4f1cda228..03330fb4a3c73dfb06d335eee540c30a7d367485 100644 --- a/seta_v2/trace_stats/README.md +++ b/seta_v2/trace_stats/README.md @@ -1,49 +1,64 @@ -# Trajectory statistics +# SETA v2 trace statistics -112,008 attempts across 3,107 tasks of SETA v2, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3-30B-A3B-Instruct-2507 | Qwen/Qwen3-30B-A3B (thinking) | -| --- | ---: | ---: | -| Recorded attempts | 56,010 | 55,998 | -| Median agent turns | 12 | 18 | -| Median input tokens per attempt | 76,716 | 150,250 | -| Median output tokens per attempt | 6,343 | 25,012 | -| Median agent runtime (seconds) | 146 | 477 | -| Median verifier runtime (seconds) | 5 | 5 | +Of 112,008 recorded attempts, 111,700 have saved conversations and 308 do not. 5,789 attempts have no raw verifier reward. Missing measurements remain missing. These counts cover all combined traces, including 12,584 historical attempts outside the pass@16 evaluation. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -![Trace measurements](plots/all-models-trace-distributions.png) +Token totals sum calls across an attempt and are not the context window. -![Trace outcomes](plots/all-models-trace-outcomes.png) +Measurements cover 3,107 tasks, including kept and unselected tasks, with each attempt counted once. -![Most frequent commands](plots/all-models-trace-commands.png) +## Evaluation settings -Commands are counted as submitted, not as successful executions. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B | Terminus-2 | On | 16 | 0.6 | 0.95 | 20 | 32,768 | 8,192 | min_p=0 | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 16 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | -Per-model plots: [Qwen/Qwen3-30B-A3B-Instruct-2507 commands](plots/qwen3-30b-instruct-2507-trace-commands.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 distributions](plots/qwen3-30b-instruct-2507-trace-distributions.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 outcomes](plots/qwen3-30b-instruct-2507-trace-outcomes.png), [Qwen/Qwen3-30B-A3B (thinking) commands](plots/qwen3-30b-thinking-trace-commands.png), [Qwen/Qwen3-30B-A3B (thinking) distributions](plots/qwen3-30b-thinking-trace-distributions.png), [Qwen/Qwen3-30B-A3B (thinking) outcomes](plots/qwen3-30b-thinking-trace-outcomes.png). +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Data +**Time limits:** agent 3,600 seconds; verifier 120, 900 seconds (sampled configurations; limits vary by task where multiple values are shown). The broader saved setup reports task-specific agent limits of 180 to 3,600 seconds and verifier limits of 120 to 1,800 seconds. -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +**Output-limit evidence:** The 7 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -Attempts without a saved conversation: 308; without a verifier reward: 5,789. Their missing measurements stay empty. +## Measurements -## Evaluation settings +### Qwen/Qwen3-30B-A3B-Instruct-2507 + +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 55,864 | 12 | 88 | +| Input tokens per attempt | 55,864 | 76,716 | 1,519,991 | +| Output tokens per attempt | 55,864 | 6,343 | 94,834 | +| Largest recorded step, prompt plus output tokens | 55,718 | 13,144 | 27,313 | +| Agent runtime (seconds) | 55,864 | 145.55 | 2,060.53 | +| Verifier runtime (seconds) | 53,246 | 4.69 | 9.52 | +| Whole-attempt runtime (seconds) | 56,010 | 184.99 | 2,104.03 | -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +### Qwen/Qwen3-30B-A3B -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | on | {"min_p": 0, "temperature": 0.6, "top_k": 20, "top_p": 0.95} | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 55,836 | 18 | 163 | +| Input tokens per attempt | 55,836 | 150,250 | 2,561,426 | +| Output tokens per attempt | 55,836 | 25,012 | 215,224 | +| Largest recorded step, prompt plus output tokens | 55,730 | 17,184 | 26,295 | +| Agent runtime (seconds) | 55,836 | 477.32 | 3,600 | +| Verifier runtime (seconds) | 53,263 | 4.65 | 8.96 | +| Whole-attempt runtime (seconds) | 55,998 | 518.31 | 3,633.18 | -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -The table reports settings found in **7 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](../metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +## Plots + +![Trace measurements](plots/all-models-trace-distributions.png) + +![Trace outcomes](plots/all-models-trace-outcomes.png) + +![Most frequent commands](plots/all-models-trace-commands.png) diff --git a/seta_v3/README.md b/seta_v3/README.md index 0c40ac80e3bdd6a542f75f847ddb083b18ec5064..0e693cdc582b7d21081104ef369087e57df9ee51 100644 --- a/seta_v3/README.md +++ b/seta_v3/README.md @@ -2,49 +2,64 @@ ## Dataset stats -- **Tasks:** 861 tasks had no success in v2; all 861 were evaluated for v3; **497 newly kept**, **364 still without success**. -- **Qwen/Qwen3.8-27B (thinking off):** 434 of those tasks, 2 attempts each; **221 solved**, **213 unsolved**, **pass@2 50.92%**. The other 427 tasks had not yet been evaluated by this model. -- **Qwen/Qwen3.8-27B (thinking on):** all 213 thinking-off failures plus the 427 previously untested tasks, **640 tasks**, 2 main attempts each; **275 solved at pass@2 42.97%**, plus **1 additional retry success**. That gives **276 newly solved** and **364 still unsolved** in this cohort. -- **Thinking-on breakdown:** of the 213 thinking-off failures, **37 solved in the main run** (pass@2 **17.37%**) and **1 more in retry**, leaving 175 unsolved. Of the 427 previously untested tasks, **238 solved** (pass@2 **55.74%**), leaving 189 unsolved. -- **Additional evaluation coverage:** 12 tasks already kept in v2 were also evaluated with thinking off; 26 thinking-off successes on earlier task instructions were rechecked with thinking on. These do not add to the 861-task denominator. +- **Tasks:** 861 had no successful attempt in v2; all 861 evaluated for v3; 497 newly kept; 364 still without success. +- **Teacher evaluation:** Qwen/Qwen3.8-27B with thinking off and thinking on, 2 main attempts per task per configuration (pass@2); thinking-on retries are reported separately. +- **Thinking off:** 221/434 solved. +- **Thinking on:** 38/213 earlier failures solved, including one retry, plus 238/427 previously untested tasks. +- **Additional coverage:** 12 tasks already kept in v2 evaluated with thinking off; 26 earlier-instruction successes rechecked with thinking on. These do not add to the 861-task denominator. - **Recorded attempts:** 2,276 new attempts (892 thinking off; 1,332 main thinking on; 52 thinking-on retries). - **Saved conversations:** 2,260 (891 thinking off; 1,369 thinking on). -`tasks.parquet` contains only the 497 newly kept tasks absent from v2. A task is kept after at least one verifier-successful attempt in these evaluations, including the separate retries. Earlier kept tasks remain in v2. These tasks were harder for the earlier evaluation; that does not establish intrinsic difficulty for every model. +`tasks.parquet` contains only the 497 newly kept tasks absent from v2. We keep tasks with at least one verifier-successful attempt in these evaluations, including the separate retries. The 2,246 tasks already kept in v2 remain separate. ## Evaluation settings -| Model | Agent | Thinking | Main attempts/task | Temperature | Top-p | Top-k | Min-p | Presence penalty | Repetition penalty | Context tokens | Intended output tokens/call | -| --- | --- | --- | ---: | ---: | ---: | ---: | --- | ---: | ---: | ---: | ---: | -| Qwen/Qwen3.8-27B | Terminus 2 | off | 2 | 0.7 | 0.8 | 20 | Not recorded | 1.5 | 1.0 | 32,768 | 8,192 | -| Qwen/Qwen3.8-27B | Terminus 2 | on | 2 | 1.0 | 0.95 | 20 | 0 | 0 | 1.0 | 32,768 | 8,192 | +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | Off | 2 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | presence_penalty=1.5; repetition_penalty=1 | +| Qwen/Qwen3.8-27B | Terminus-2 | On | 2 | 1 | 0.95 | 20 | 32,768 | 8,192 | min_p=0; presence_penalty=0; repetition_penalty=1 | -The thinking-off run did not forward its intended output cap; the actual backend fallback cap is unknown. All 1,384 thinking-on trial configurations forward `llm_call_kwargs.max_tokens=8192`. Context covers prompt plus response for one call, including thinking. Agent limits are task-specific: 360 to 3,600 seconds with thinking off and 360 to 3,600 seconds with thinking on. Verifier limits remain task-specific; no complete per-configuration range was audited here. Saved contracts and settings audits are in `metadata/`. +Context includes prompt and response for one model call. Output includes thinking tokens; neither is a whole-attempt token total. Main runs use two attempts per task; the thinking-on retry adds two attempts on 26 tasks. + +**Thinking off:** All 892 saved thinking-off trial configurations use a top-level output limit that was not forwarded as an explicit request cap; the actual fallback cap is unknown. Agent limits are 360, 600, 900 or 3,600 seconds by task; verifier limits are 120, 300, 360, 600 or 900 seconds. Min-p is not recorded. + +**Thinking on:** All 1,384 saved thinking-on trial configurations contain an explicit 8,192-token request cap. HTTP payloads and server enforcement were not audited. Agent limits are task-specific, from 360 to 3,600 seconds. Verifier limits remain task-specific; a complete range has not been audited. The configurations cover different task subsets, so their scores are not a controlled comparison of thinking modes. Thinking off ran stages 6 and 7 only, without stage-1 validation or path evidence. Thinking on ran stage 1 on all 666 evaluated tasks: 659 passed, 5 had setup errors, 1 failed the schema check and 1 could not complete repository-leak checks. Relative-path evidence was collected for 339 of 342 tasks that used relative instruction paths. These failures remain failures or incomplete checks; teacher success does not establish that all validation checks passed. Thinking on requested isolated networking but fell back to host networking; both configurations had internet access. ## Results -Main-run scores use exactly two attempts per task. Terminal errors count as unsuccessful under `terminal-errors-zero-v2`; their raw verifier rewards remain null. Retry attempts are preserved separately and excluded from pass@2. +Main-run results use exactly two attempts per task. Terminal errors count as unsuccessful under the recorded policy while raw missing verifier grades remain null. Retries are separate and excluded from pass@2. Thinking off and thinking on evaluated different task sets, so these are not controlled comparisons of thinking modes. -| Model and thinking mode | Evaluated subset | Tasks | Solved in main run | Pass@1 | Pass@2 | -| --- | --- | ---: | ---: | ---: | ---: | -| Qwen/Qwen3.8-27B (off) | v2-unsolved | 434 | 221 | 41.94% | 50.92% | -| Qwen/Qwen3.8-27B (on) | Remaining after thinking off | 640 | 275 | 34.53% | 42.97% | -| Qwen/Qwen3.8-27B (on) | Of those: thinking-off failures | 213 | 37 | 11.27% | 17.37% | -| Qwen/Qwen3.8-27B (on) | Of those: previously untested | 427 | 238 | 46.14% | 55.74% | -| Qwen/Qwen3.8-27B (off) | Already kept in v2 | 12 | 11 | 79.17% | 91.67% | -| Qwen/Qwen3.8-27B (on) | Earlier-instruction successes rechecked | 26 | 21 | 73.08% | 80.77% | +| Successful main attempts | Qwen/Qwen3.8-27B (thinking off: all evaluated, 446 tasks) | Qwen/Qwen3.8-27B (thinking on: all evaluated, 666 tasks) | +| --- | ---: | ---: | +| 0/2 | 214 | 370 | +| 1/2 | 81 | 112 | +| 2/2 | 151 | 184 | -The 213-task and 427-task rows partition the 640-task row. The retry evaluated 26 tasks with terminal errors and no main-run success, adding one success in the 213-task subset. Across thinking off, thinking on and retries, 497 of the original 861 tasks were solved (57.72%); this mixed-budget fraction is not a pass@2 score. +Breakdown by evaluated subset. The two thinking-off columns partition its 446 tasks; the three thinking-on columns partition its 666 tasks. -| Successful main attempts | Thinking off: 434 v2-unsolved | Thinking on: 213 earlier failures | Thinking on: 427 previously untested | Thinking off: 12 v2-kept | Thinking on: 26 rechecks | +| Successful main attempts | Thinking off: v2-unsolved (434) | Thinking off: already kept in v2 (12) | Thinking on: earlier failures (213) | Thinking on: previously untested (427) | Thinking on: earlier-instruction rechecks (26) | | --- | ---: | ---: | ---: | ---: | ---: | -| 0/2 | 213 | 176 | 189 | 1 | 5 | -| 1/2 | 78 | 26 | 82 | 3 | 4 | -| 2/2 | 143 | 11 | 156 | 8 | 17 | +| 0/2 | 213 | 1 | 176 | 189 | 5 | +| 1/2 | 78 | 3 | 26 | 82 | 4 | +| 2/2 | 143 | 8 | 11 | 156 | 17 | + +| Model | Thinking mode and subset | Tasks | Pass@1 | Pass@2 | +| --- | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Thinking off: all evaluated | 446 | 42.94% | 52.02% | +| Qwen/Qwen3.8-27B | Thinking off: v2-unsolved | 434 | 41.94% | 50.92% | +| Qwen/Qwen3.8-27B | Thinking off: already kept in v2 | 12 | 79.17% | 91.67% | +| Qwen/Qwen3.8-27B | Thinking on: all evaluated | 666 | 36.04% | 44.44% | +| Qwen/Qwen3.8-27B | Thinking on: earlier failures | 213 | 11.27% | 17.37% | +| Qwen/Qwen3.8-27B | Thinking on: previously untested | 427 | 46.14% | 55.74% | +| Qwen/Qwen3.8-27B | Thinking on: earlier-instruction rechecks | 26 | 73.08% | 80.77% | + +Among the 640 tasks remaining after thinking off, thinking on solved 275 in its main run (pass@2 42.97%): 37/213 earlier failures and 238/427 previously untested tasks. One additional retry success raises the first count to 38/213 and the thinking-on addition to 276 tasks. Across both configurations and retries, 497/861 tasks were solved; this mixed-budget fraction is not a pass@2 score. + +Pass@1 is estimated from the two recorded attempts per task; pass@2 is the fraction with at least one main-run success. -![Pass@1 and pass@2 by evaluation cohort](metadata/plots/passk.png) +![Pass rates by thinking mode and evaluated subset, excluding retries](plots/readme-pass-at-k.png) ## Files and usage diff --git a/seta_v3/plots/readme-pass-at-k.png b/seta_v3/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..21c8390652d96844cb7f1aaf4773328ac753f4c3 --- /dev/null +++ b/seta_v3/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5207fa9d7c3b3a018bf1ae13361d0128700a2f334275ec514af6a127dd9f1f5 +size 95368 diff --git a/seta_v3/trace_stats/README.md b/seta_v3/trace_stats/README.md index c2afb258fb691e72d656a61d20b0cdf19ea64a03..229e92cb04158dbd7261fa585d344942a5330364 100644 --- a/seta_v3/trace_stats/README.md +++ b/seta_v3/trace_stats/README.md @@ -6,12 +6,16 @@ Only the new Qwen/Qwen3.8-27B evaluations are included, with each attempt counte ## Evaluation settings -| Model | Agent | Thinking | Main attempts/task | Temperature | Top-p | Top-k | Min-p | Presence penalty | Repetition penalty | Context tokens | Intended output tokens/call | -| --- | --- | --- | ---: | ---: | ---: | ---: | --- | ---: | ---: | ---: | ---: | -| Qwen/Qwen3.8-27B | Terminus 2 | off | 2 | 0.7 | 0.8 | 20 | Not recorded | 1.5 | 1.0 | 32,768 | 8,192 | -| Qwen/Qwen3.8-27B | Terminus 2 | on | 2 | 1.0 | 0.95 | 20 | 0 | 0 | 1.0 | 32,768 | 8,192 | +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | Off | 2 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | presence_penalty=1.5; repetition_penalty=1 | +| Qwen/Qwen3.8-27B | Terminus-2 | On | 2 | 1 | 0.95 | 20 | 32,768 | 8,192 | min_p=0; presence_penalty=0; repetition_penalty=1 | -The thinking-off run did not forward its intended output cap; the actual backend fallback cap is unknown. All 1,384 thinking-on trial configurations forward `llm_call_kwargs.max_tokens=8192`. Context covers prompt plus response for one call, including thinking. Agent limits are task-specific: 360 to 3,600 seconds with thinking off and 360 to 3,600 seconds with thinking on. Verifier limits remain task-specific; no complete per-configuration range was audited here. Saved contracts and settings audits are in `metadata/`. +Context includes prompt and response for one model call. Output includes thinking tokens; neither is a whole-attempt token total. Main runs use two attempts per task; the thinking-on retry adds two attempts on 26 tasks. + +**Thinking off:** All 892 saved thinking-off trial configurations use a top-level output limit that was not forwarded as an explicit request cap; the actual fallback cap is unknown. Agent limits are 360, 600, 900 or 3,600 seconds by task; verifier limits are 120, 300, 360, 600 or 900 seconds. Min-p is not recorded. + +**Thinking on:** All 1,384 saved thinking-on trial configurations contain an explicit 8,192-token request cap. HTTP payloads and server enforcement were not audited. Agent limits are task-specific, from 360 to 3,600 seconds. Verifier limits remain task-specific; a complete range has not been audited. ## Measurements diff --git a/seta_v3/trace_stats/qwen38-27b-thinking-off/README.md b/seta_v3/trace_stats/qwen38-27b-thinking-off/README.md index ff3d5f8ac5d1bf7b6ad3c0240080534f08384de6..8860d6c30f8923c1faf339522f529d420b1959b8 100644 --- a/seta_v3/trace_stats/qwen38-27b-thinking-off/README.md +++ b/seta_v3/trace_stats/qwen38-27b-thinking-off/README.md @@ -2,15 +2,17 @@ ## Coverage -This directory covers only the thinking-off evaluation, including unsuccessful attempts. Coverage and missing measurements are shown below. No quality filtering was applied. +Only the new thinking-off evaluation is included: 446 tasks, 892 attempts, 891 saved conversations, 1 missing conversations and 7 missing raw verifier grades. Successful and unsuccessful attempts are included; no quality filtering was applied. ## Evaluation settings -| Model | Agent | Thinking | Main attempts/task | Temperature | Top-p | Top-k | Min-p | Presence penalty | Repetition penalty | Context tokens | Intended output tokens/call | -| --- | --- | --- | ---: | ---: | ---: | ---: | --- | ---: | ---: | ---: | ---: | -| Qwen/Qwen3.8-27B | Terminus 2 | off | 2 | 0.7 | 0.8 | 20 | Not recorded | 1.5 | 1.0 | 32,768 | 8,192 | +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | Off | 2 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | presence_penalty=1.5; repetition_penalty=1 | -The thinking-off run did not forward its intended output cap; the actual backend fallback cap is unknown. All 1,384 thinking-on trial configurations forward `llm_call_kwargs.max_tokens=8192`. Context covers prompt plus response for one call, including thinking. Agent limits are task-specific: 360 to 3,600 seconds with thinking off and 360 to 3,600 seconds with thinking on. Verifier limits remain task-specific; no complete per-configuration range was audited here. Saved contracts and settings audits are in `metadata/`. +Context includes prompt and response for one model call. Output includes thinking tokens; neither is a whole-attempt token total. Main runs use two attempts per task; the thinking-on retry adds two attempts on 26 tasks. + +All 892 saved thinking-off trial configurations use a top-level output limit that was not forwarded as an explicit request cap; the actual fallback cap is unknown. Agent limits are 360, 600, 900 or 3,600 seconds by task; verifier limits are 120, 300, 360, 600 or 900 seconds. Min-p is not recorded. ## Measurements @@ -35,4 +37,3 @@ Medians use recorded measurements; missing values remain empty. Token totals spa ![Trace outcomes](plots/all-models-trace-outcomes.png) ![Submitted commands](plots/all-models-trace-commands.png) - diff --git a/seta_v3/trace_stats/qwen38-27b-thinking-on/README.md b/seta_v3/trace_stats/qwen38-27b-thinking-on/README.md index ad69b7849cd8659f8b45748b1b3289c051a5932e..c79326bf6aeac906c64e0ecdf05dcc03e130c961 100644 --- a/seta_v3/trace_stats/qwen38-27b-thinking-on/README.md +++ b/seta_v3/trace_stats/qwen38-27b-thinking-on/README.md @@ -2,15 +2,17 @@ ## Coverage -This directory covers only the thinking-on evaluation, including unsuccessful attempts. Coverage and missing measurements are shown below. No quality filtering was applied. +Only the new thinking-on evaluation is included: 666 tasks, 1,384 attempts, 1,369 saved conversations, 15 missing conversations and 71 missing raw verifier grades. Successful and unsuccessful attempts are included; no quality filtering was applied. The statistics include 52 separate retry attempts; the main pass@2 results exclude them. ## Evaluation settings -| Model | Agent | Thinking | Main attempts/task | Temperature | Top-p | Top-k | Min-p | Presence penalty | Repetition penalty | Context tokens | Intended output tokens/call | -| --- | --- | --- | ---: | ---: | ---: | ---: | --- | ---: | ---: | ---: | ---: | -| Qwen/Qwen3.8-27B | Terminus 2 | on | 2 | 1.0 | 0.95 | 20 | 0 | 0 | 1.0 | 32,768 | 8,192 | +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | On | 2 | 1 | 0.95 | 20 | 32,768 | 8,192 | min_p=0; presence_penalty=0; repetition_penalty=1 | -The thinking-off run did not forward its intended output cap; the actual backend fallback cap is unknown. All 1,384 thinking-on trial configurations forward `llm_call_kwargs.max_tokens=8192`. Context covers prompt plus response for one call, including thinking. Agent limits are task-specific: 360 to 3,600 seconds with thinking off and 360 to 3,600 seconds with thinking on. Verifier limits remain task-specific; no complete per-configuration range was audited here. Saved contracts and settings audits are in `metadata/`. +Context includes prompt and response for one model call. Output includes thinking tokens; neither is a whole-attempt token total. Main runs use two attempts per task; the thinking-on retry adds two attempts on 26 tasks. + +All 1,384 saved thinking-on trial configurations contain an explicit 8,192-token request cap. HTTP payloads and server enforcement were not audited. Agent limits are task-specific, from 360 to 3,600 seconds. Verifier limits remain task-specific; a complete range has not been audited. ## Measurements @@ -35,4 +37,3 @@ Medians use recorded measurements; missing values remain empty. Token totals spa ![Trace outcomes](plots/all-models-trace-outcomes.png) ![Submitted commands](plots/all-models-trace-commands.png) - diff --git a/swelego_v2/README.md b/swelego_v2/README.md index 6fe46571ead4c6aff4a517643c6a73b3ab5d941c..d74578fa669c6b1bcca5c2797cde44f5f1641f40 100644 --- a/swelego_v2/README.md +++ b/swelego_v2/README.md @@ -2,90 +2,80 @@ ## Dataset stats -- **Evaluated tasks:** 4,277 -- **Teacher:** `Qwen/Qwen3-30B-A3B-Instruct-2507` (thinking off) -- **Evaluation budget:** pass@8, eight attempts per task -- **Kept:** 1,245 / 4,277 tasks, **29.11% pass@8** -- **No recorded success:** 3,032 tasks -- **Saved conversations:** 32,803 / 34,216 attempts +- **Tasks:** 4,277 in v1; 1,245 kept in v2. +- **Teacher evaluation:** Qwen/Qwen3-30B-A3B-Instruct-2507 (thinking off), 8 attempts per task (pass@8). +- **Coverage:** 4,277/4,277 tasks evaluated; 34,216 evaluation attempts. +- **Saved conversations:** 32,803 of 34,216 published attempt records, including successful and unsuccessful attempts. -## Dataset characteristics +We keep tasks with at least one verifier-successful attempt from any evaluated model. -SWE-Lego contains repository-level software-repair tasks. This release selects tasks from [v1](../swelego_v1/README.md) that the teacher solved at least once. Task contents are preserved; selection decisions and pinned source provenance are in [the selection metadata](metadata/selection.json). +## Evaluation settings -## Evaluation setup and coverage +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 8 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | -`Qwen/Qwen3-30B-A3B-Instruct-2507` (thinking off) ran with **Terminus-2**, with eight classified outcomes for every task. Recorded terminal errors count as unsuccessful attempts; complete outcome coverage does not mean every attempt reached the agent or verifier. +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -| Setting | Recorded configuration | -| --- | --- | -| Context window per model call | 32,768 tokens, prompt plus response | -| Intended output limit per model call | 8,192 tokens | -| Sampling | Temperature 0.7, top-p 0.8, top-k 20, min-p 0 | -| Agent timeout | 1,200 seconds in sampled resolved configurations | -| Verifier timeout | 1,800 seconds in sampled resolved configurations | +**Time limits:** agent 1,200 seconds; verifier 1,800 seconds (sampled configurations; limits vary by task where multiple values are shown). -**The intended output cap was not forwarded to the API in the sampled runs.** Their actual fallback cap is unknown. These settings come from six saved trial configurations, not an exhaustive audit; historical runs and retries may differ. See [settings and evidence](metadata/evaluation-settings-audit.json). Token budgets apply per model call, not across an entire attempt. +**Output-limit evidence:** The 6 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -## Pass@k and success distribution +## Results -Results for `Qwen/Qwen3-30B-A3B-Instruct-2507` (thinking off), over **4,277 tasks and 34,216 attempts**: +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. -| Metric | Score | +| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 (4,277 tasks) | | --- | ---: | -| Pass@1 | 10.73% | -| Pass@4 | 22.25% | -| Pass@8 | 29.11% | +| 0/8 | 3,032 | +| 1/8 | 468 | +| 2/8 | 221 | +| 3/8 | 144 | +| 4/8 | 133 | +| 5/8 | 77 | +| 6/8 | 72 | +| 7/8 | 58 | +| 8/8 | 72 | -Pass@1 and pass@4 use unbiased estimates from the eight attempts. Pass@8 is the observed fraction of tasks solved at least once. Infrastructure failures are included as unsuccessful attempts, so these scores are lower bounds for the model. +| Model | Evaluated tasks | Pass@1 | Pass@2 | Pass@4 | Pass@8 | +| --- | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | 4,277 | 10.73% | 16.01% | 22.25% | 29.11% | -| Successful attempts out of 8 | Tasks | -| --- | ---: | -| 0 | 3,032 | -| 1 | 468 | -| 2 | 221 | -| 3 | 144 | -| 4 | 133 | -| 5 | 77 | -| 6 | 72 | -| 7 | 58 | -| 8 | 72 | +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. -Across the recorded samples, 72 tasks were always solved, 3,032 had no success, and 1,173 had mixed outcomes. +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -## Trajectories and limitations +## Files and usage -The [traces](traces) include kept and unselected tasks, with raw verifier rewards and classified outcomes kept separate. Of 1,413 attempts without a saved conversation, 1,412 ended before or during agent start; one lacks its source archive. All 1,580 attempts without a raw verifier grade are recorded terminal errors scored zero, rather than unknown outcomes. +`tasks.parquet` contains 1,245 kept tasks. -Recorded terminations include 4,060 task timeouts, 24 task memory-limit failures and 29 verifier errors. See [trace statistics and plots](trace_stats/README.md) for the full breakdown and measurement coverage. +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. -Traces for 31 kept tasks include earlier task versions. The repository-history and timestamp leak checks introduced after the original validation remain incomplete for this release; teacher outcomes do not establish that those checks passed. +Success-count columns: `instruct_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. -## Runtime and token usage +```python +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -Measured attempt distributions for `Qwen/Qwen3-30B-A3B-Instruct-2507` (thinking off): +path = hf_hub_download( + "FWeindel/validated-tasks", + "swelego_v2/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("instruct_successes", "=", 1)]) +``` -| Metric | Attempts measured | Median | 90th percentile | -| --- | ---: | ---: | ---: | -| Agent turns | 32,804 | 15 | 102 | -| Input tokens per attempt | 32,804 | 98,424 | 1,649,575 | -| Output tokens per attempt | 32,804 | 4,885 | 75,449 | -| Peak context tokens | 32,803 | 12,380 | 25,606 | -| Agent runtime, seconds | 32,804 | 100 | 1,200 | -| Verifier runtime, seconds | 32,733 | 7 | 15 | -| Trial runtime, seconds | 34,216 | 166 | 1,232 | +31 kept tasks have traces from earlier task versions; the published file preserves the pinned v1 payloads. -Missing measurements are excluded, not replaced with zero. Input/output totals span all calls in an attempt. The [measurement summary](trace_stats/data/summary.json) and [trace provenance](trace_stats/data/provenance.json) contain detailed definitions and evidence. +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. -The Helma rerun used Apptainer 1.5.4, Python 3.12.14 and Harbor 0.8.1, with [Marin Harbor](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) and [runtime patches](https://github.com/MLI-lab/OT-next-data/tree/main/harbor_patches). Its frozen configuration is preserved in [the run reports](metadata/rerun-reports-20261020.tar.gz). +## Trace statistics -## Reproduce the selection +Of 34,216 recorded attempts, 32,803 have saved conversations and 1,413 do not. 1,580 attempts have no raw verifier reward. Missing measurements remain missing. Of the missing conversations, 1,412 were never produced because setup or agent start failed; one attempt has a verifier grade but no available source conversation. -`patch.py` rebuilds `tasks.parquet` from the pinned v1 source and [selection metadata](metadata/selection.json). For example, to select kept tasks with exactly one successful attempt: +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -```python -import pyarrow.parquet as pq -subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)]) -``` +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -`instruct_graded_attempts` counts classified outcomes, including terminal errors scored zero. `instruct_verifier_graded_attempts` counts actual verifier grades; `instruct_missing_attempts` counts unknown outcomes. +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/swelego_v2/plots/readme-pass-at-k.png b/swelego_v2/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..acb0eee004b4e9d61259d8177d8c0d7d57ba1628 --- /dev/null +++ b/swelego_v2/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43710fb4197c7c9892a47a0151f331180b8c19be9005f67c97130e469762fe2b +size 43409 diff --git a/swelego_v2/trace_stats/README.md b/swelego_v2/trace_stats/README.md index 75011344350872ee4dfae134506621a38ba094cf..ff7b45dbd8992b73dee51def3ab6dee777bbcd58 100644 --- a/swelego_v2/trace_stats/README.md +++ b/swelego_v2/trace_stats/README.md @@ -1,51 +1,51 @@ -# Trajectory statistics +# SWE-Lego v2 trace statistics -34,216 attempts across 4,277 tasks of SWE LEGO v2, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3-30B-A3B-Instruct-2507 | -| --- | ---: | -| Recorded attempts | 34,216 | -| Median agent turns | 15 | -| Median input tokens per attempt | 98,424 | -| Median output tokens per attempt | 4,885 | -| Median agent runtime (seconds) | 100 | -| Median verifier runtime (seconds) | 7 | +Of 34,216 recorded attempts, 32,803 have saved conversations and 1,413 do not. 1,580 attempts have no raw verifier reward. Missing measurements remain missing. Of the missing conversations, 1,412 were never produced because setup or agent start failed; one attempt has a verifier grade but no available source conversation. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -![Trace measurements](plots/all-models-trace-distributions.png) +Token totals sum calls across an attempt and are not the context window. -![Trace outcomes](plots/all-models-trace-outcomes.png) +Measurements cover 4,277 tasks, including kept and unselected tasks, with each attempt counted once. -![Most frequent commands](plots/all-models-trace-commands.png) +## Evaluation settings -Commands are counted as submitted, not as successful executions. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 8 | 0.7 | 0.8 | 20 | 32,768 | 8,192 | min_p=0 | -Per-model plots: [Qwen/Qwen3-30B-A3B-Instruct-2507 commands](plots/instruct-trace-commands.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 distributions](plots/instruct-trace-distributions.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 outcomes](plots/instruct-trace-outcomes.png). +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Data +**Time limits:** agent 1,200 seconds; verifier 1,800 seconds (sampled configurations; limits vary by task where multiple values are shown). -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +**Output-limit evidence:** The 6 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -Attempts without a saved conversation: 1,413; without a verifier reward: 1,580. Their missing measurements stay empty. +## Measurements +### Qwen/Qwen3-30B-A3B-Instruct-2507 -Evaluation budgets: see the [dataset evaluation setup](../README.md). Per-call context and output-limit enforcement are not established by token-usage statistics. +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 32,804 | 15 | 102 | +| Input tokens per attempt | 32,804 | 98,424 | 1,649,575 | +| Output tokens per attempt | 32,804 | 4,885 | 75,449 | +| Largest recorded step, prompt plus output tokens | 32,803 | 12,380 | 25,606 | +| Agent runtime (seconds) | 32,804 | 99.84 | 1,200 | +| Verifier runtime (seconds) | 32,733 | 7.46 | 14.51 | +| Whole-attempt runtime (seconds) | 34,216 | 165.76 | 1,232.19 | -## Evaluation settings +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 32768 | 8192 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +## Plots -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +![Trace measurements](plots/all-models-trace-distributions.png) -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +![Trace outcomes](plots/all-models-trace-outcomes.png) -The table reports settings found in **6 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](../metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +![Most frequent commands](plots/all-models-trace-commands.png) diff --git a/termigen_v2/README.md b/termigen_v2/README.md index 7975278ca5b31030124adf990a2952d71411cc20..8e68caddf3bf6ff43b8d41cba342cbbb9498f7bb 100644 --- a/termigen_v2/README.md +++ b/termigen_v2/README.md @@ -2,78 +2,76 @@ ## Dataset stats -- **v1 tasks evaluated:** 3,336 -- **Kept:** 2,277 (solved at least once by either model) -- **Excluded:** 1,059 (no recorded success) -- **Attempts:** pass@4 with Qwen/Qwen3-Coder-30B-A3B-Instruct and Qwen/Qwen3-30B-A3B-Instruct-2507 (26,688 in total) +- **Tasks:** 3,336 in v1; 2,277 kept in v2. +- **Teacher evaluation:** Qwen/Qwen3-30B-A3B-Instruct-2507 (thinking off), 4 attempts per task (pass@4); Qwen/Qwen3-Coder-30B-A3B-Instruct (thinking off), 4 attempts per task (pass@4). +- **Coverage:** 3,336/3,336 tasks evaluated; 26,688 evaluation attempts. +- **Saved conversations:** 26,624 of 26,688 published attempt records, including successful and unsuccessful attempts. -## Evaluation settings - -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. - -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-Coder-30B-A3B-Instruct | 65536 | 16384 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"repetition_penalty": 1.05, "temperature": 0.7, "top_k": 20, "top_p": 0.95} | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 65536 | 16384 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +We keep tasks with at least one verifier-successful attempt from any evaluated model. -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +## Evaluation settings -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-Coder-30B-A3B-Instruct | Terminus-2 | Off | 4 | 0.7 | 0.95 | 20 | 65,536 | 16,384 | repetition_penalty=1.05 | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 4 | 0.7 | 0.8 | 20 | 65,536 | 16,384 | min_p=0 | -The table reports settings found in **4 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Pass@4 filtering +**Time limits:** agent 900 seconds; verifier 240 seconds (sampled configurations; limits vary by task where multiple values are shown). -| Model | Tasks evaluated | Attempts | Tasks solved at least once | -| --- | ---: | ---: | ---: | -| Qwen/Qwen3-Coder-30B-A3B-Instruct | 3,336 | 13,344 | 2,123 | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 3,336 | 13,344 | 1,850 | -| Either model | 3,336 | 26,688 | 2,277 | +**Output-limit evidence:** The 4 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -Every evaluated task has 4 attempts per model; 142 Qwen/Qwen3-Coder-30B-A3B-Instruct and 179 Qwen/Qwen3-30B-A3B-Instruct-2507 attempts have no recorded reward and count as unknown, not as failures. +## Results -### Successful attempts among kept tasks +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. -| Successful attempts | Qwen/Qwen3-Coder-30B-A3B-Instruct tasks | Qwen/Qwen3-30B-A3B-Instruct-2507 tasks | +| Successful attempts | Qwen/Qwen3-30B-A3B-Instruct-2507 (3,336 tasks) | Qwen/Qwen3-Coder-30B-A3B-Instruct (3,336 tasks) | | --- | ---: | ---: | -| 0/4 | 154 | 427 | -| 1/4 | 302 | 417 | -| 2/4 | 338 | 359 | -| 3/4 | 444 | 421 | -| 4/4 | 1,039 | 653 | +| 0/4 | 1,486 | 1,213 | +| 1/4 | 417 | 302 | +| 2/4 | 359 | 338 | +| 3/4 | 421 | 444 | +| 4/4 | 653 | 1,039 | -## Select tasks by successful-attempt count +| Model | Evaluated tasks | Pass@1 | Pass@2 | Pass@4 | +| --- | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | 3,336 | 37.54% | 47.41% | 55.46% | +| Qwen/Qwen3-Coder-30B-A3B-Instruct | 3,336 | 48.46% | 57.42% | 63.64% | -You can select tasks by how often they were solved, for example all tasks Qwen/Qwen3-30B-A3B-Instruct-2507 solved exactly once: +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. There are 142 attempts without a recorded reward for Qwen/Qwen3-Coder-30B-A3B-Instruct and 179 for Qwen/Qwen3-30B-A3B-Instruct-2507. These remain unknown, not verified failures. The plotted values are observed-success lower bounds with unknown outcomes unresolved, not complete-outcome pass-rate estimates. -```python -import pyarrow.parquet as pq +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -subset = pq.read_table("tasks.parquet", filters=[("instruct_successes", "=", 1)]) -``` +## Files and usage -## Traces +`tasks.parquet` contains 2,277 kept tasks. -`traces/` holds all 26,688 recorded attempts, successful and failed, from every evaluated task, one row per attempt (Zstandard-compressed Parquet). Columns include `task_id`, `model`, `attempt_id`, `reward`, `verifier_reward`, `retained_v2` and the full `trajectory` and `result` JSON. 64 attempts have no saved conversation (`trajectory_found=false`). +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. + +Success-count columns: `instruct_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. ```python -import json -import pyarrow.dataset as ds +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -traces = ds.dataset("traces", format="parquet") -failed = traces.to_table(filter=ds.field("reward") == 0) -conversation = json.loads(failed["trajectory"][0].as_py()) +path = hf_hub_download( + "FWeindel/validated-tasks", + "termigen_v2/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("instruct_successes", "=", 1)]) ``` -Trajectory statistics (plots, command distribution) are in [trace_stats](trace_stats/README.md). +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. -Teacher results and trajectories come from [PR #20](https://huggingface.co/datasets/FWeindel/validated-tasks/discussions/20); [recorded model results](https://huggingface.co/datasets/FWeindel/validated-tasks/discussions/11). +## Trace statistics -Not yet checked with the repository-leak checks added on 2026-10-17 (leftover git history and file timestamps that could reveal the solution). +Of 26,688 recorded attempts, 26,624 have saved conversations and 64 do not. 321 attempts have no raw verifier reward. Missing measurements remain missing. -## Runtime +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -- **Versions:** Apptainer **1.5.4**, Python **3.12.7 / 3.12.14**, Harbor **0.8.1**. -- **Implementation:** [Marin Harbor `7faf878c14b6`](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) + [our runtime patches](https://github.com/MLI-lab/OT-next-data/tree/main/harbor_patches). +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -`patch.py` rebuilds `tasks.parquet` from the current v1 tasks and the recorded results, both pinned to revision `a43f5635`. +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/termigen_v2/plots/readme-pass-at-k.png b/termigen_v2/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..3f48ce90b1012dcf4a548744d3a2152ccfbe47d5 --- /dev/null +++ b/termigen_v2/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b85a19c562f49ce79284cb35c66dcbfcd2db8ee106a0f3e9c0e7a4bd371e8e1b +size 61022 diff --git a/termigen_v2/trace_stats/README.md b/termigen_v2/trace_stats/README.md index 39dfc327b8d4b836a6fc6538d3dbf86263f5633a..d1e539c65e066130934c5dfa7454dc7c61243321 100644 --- a/termigen_v2/trace_stats/README.md +++ b/termigen_v2/trace_stats/README.md @@ -1,49 +1,64 @@ -# Trajectory statistics +# TermiGen v2 trace statistics -26,688 attempts across 3,336 tasks of termigen_v2, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3-30B-A3B-Instruct-2507 | Qwen/Qwen3-Coder-30B-A3B-Instruct | -| --- | ---: | ---: | -| Recorded attempts | 13,344 | 13,344 | -| Median agent turns | 6 | 21 | -| Median input tokens per attempt | 20,686 | 132,895 | -| Median output tokens per attempt | 2,352 | 4,356 | -| Median agent runtime (seconds) | 53 | 71 | -| Median verifier runtime (seconds) | 2 | 2 | +Of 26,688 recorded attempts, 26,624 have saved conversations and 64 do not. 321 attempts have no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. -![Trace measurements](plots/all-models-trace-distributions.png) +Token totals sum calls across an attempt and are not the context window. -![Trace outcomes](plots/all-models-trace-outcomes.png) +Measurements cover 3,336 tasks, including kept and unselected tasks, with each attempt counted once. -![Most frequent commands](plots/all-models-trace-commands.png) +## Evaluation settings -Commands are counted as submitted, not as successful executions. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3-Coder-30B-A3B-Instruct | Terminus-2 | Off | 4 | 0.7 | 0.95 | 20 | 65,536 | 16,384 | repetition_penalty=1.05 | +| Qwen/Qwen3-30B-A3B-Instruct-2507 | Terminus-2 | Off | 4 | 0.7 | 0.8 | 20 | 65,536 | 16,384 | min_p=0 | -Per-model plots: [Qwen/Qwen3-30B-A3B-Instruct-2507 commands](plots/qwen3-30b-a3b-instruct-2507-trace-commands.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 distributions](plots/qwen3-30b-a3b-instruct-2507-trace-distributions.png), [Qwen/Qwen3-30B-A3B-Instruct-2507 outcomes](plots/qwen3-30b-a3b-instruct-2507-trace-outcomes.png), [Qwen/Qwen3-Coder-30B-A3B-Instruct commands](plots/qwen3-coder-30b-a3b-instruct-trace-commands.png), [Qwen/Qwen3-Coder-30B-A3B-Instruct distributions](plots/qwen3-coder-30b-a3b-instruct-trace-distributions.png), [Qwen/Qwen3-Coder-30B-A3B-Instruct outcomes](plots/qwen3-coder-30b-a3b-instruct-trace-outcomes.png). +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Data +**Time limits:** agent 900 seconds; verifier 240 seconds (sampled configurations; limits vary by task where multiple values are shown). -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +**Output-limit evidence:** The 4 sampled saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. Settings describe the sample, not an exhaustive audit of every historical attempt or retry. -Attempts without a saved conversation: 64; without a verifier reward: 321. Their missing measurements stay empty. +## Measurements -## Evaluation settings +### Qwen/Qwen3-30B-A3B-Instruct-2507 + +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 13,312 | 6 | 33 | +| Input tokens per attempt | 13,312 | 20,686 | 814,032 | +| Output tokens per attempt | 13,312 | 2,352 | 29,875 | +| Largest recorded step, prompt plus output tokens | 13,225 | 5,816 | 51,875 | +| Agent runtime (seconds) | 13,312 | 52.86 | 900 | +| Verifier runtime (seconds) | 13,185 | 1.97 | 2.61 | +| Whole-attempt runtime (seconds) | 13,344 | 62.59 | 908.8 | -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +### Qwen/Qwen3-Coder-30B-A3B-Instruct -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3-Coder-30B-A3B-Instruct | 65536 | 16384 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"repetition_penalty": 1.05, "temperature": 0.7, "top_k": 20, "top_p": 0.95} | -| Qwen/Qwen3-30B-A3B-Instruct-2507 | 65536 | 16384 | No explicit request cap in sampled saved configs; actual fallback cap unknown | off | {"min_p": 0, "temperature": 0.7, "top_k": 20, "top_p": 0.8} | +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 13,312 | 21 | 68 | +| Input tokens per attempt | 13,312 | 132,895 | 1,043,807 | +| Output tokens per attempt | 13,312 | 4,356 | 25,537 | +| Largest recorded step, prompt plus output tokens | 13,312 | 10,671 | 37,964 | +| Agent runtime (seconds) | 13,312 | 71.44 | 362.51 | +| Verifier runtime (seconds) | 13,210 | 1.8 | 2.3 | +| Whole-attempt runtime (seconds) | 13,344 | 79.56 | 389.35 | -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. Unknown historical settings are not filled from current model defaults. +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus these sampled runs did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 or 16,384 enforced tokens, or as unlimited. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -The table reports settings found in **4 sampled saved trial configurations**, not an exhaustive audit of every attempt or direct proof of server settings. Historical runs and retries may differ. Sampling values are recorded explicitly rather than inferred from today's model-card defaults. Agent and verifier wall-clock budgets remain task-specific; sampled resolved limits are in [settings and evidence](../metadata/evaluation-settings-audit.json). Context is not the sum of tokens across all turns. The current forwarding fix does not retroactively change these pass@k results. +## Plots + +![Trace measurements](plots/all-models-trace-distributions.png) + +![Trace outcomes](plots/all-models-trace-outcomes.png) + +![Most frequent commands](plots/all-models-trace-commands.png) diff --git a/termigen_v3/README.md b/termigen_v3/README.md index fa534a781ff73eade80adc9a525b126db99834c6..77729b093249dca027d9d2ddafceaa87d3003068 100644 --- a/termigen_v3/README.md +++ b/termigen_v3/README.md @@ -2,102 +2,79 @@ ## Dataset stats -- **v1 tasks:** 3,467 (pinned source revision `5f4179046b9f321b5d0f8b1c534eb6511299e39c`) -- **Kept in v3:** 230 newly solved tasks -- **Available separately in v2:** 2,277 -- **Combined v2 + v3:** 2,507 unique tasks -- **Excluded:** 829 (evaluated, no success in v2 or in the new evaluation) -- **Not evaluated:** 131 (outside the 3,336-task v2 evaluation) -- **New evaluation:** 1,059 tasks unsolved in v2, 2 attempts each, 2,118 attempts -- **New saved trajectories:** 2,118, successful and failed +- **Tasks:** 1,059 had no successful attempt in v2; all 1,059 evaluated for v3; 230 newly kept; 829 still without success. +- **Teacher evaluation:** Qwen/Qwen3.8-27B (thinking on), 2 attempts per task (pass@2). +- **Coverage:** 1,059 tasks evaluated; 2,118 evaluation attempts. +- **Saved conversations:** 2,118 of 2,118 published attempt records, including successful and unsuccessful attempts. -`tasks.parquet` contains only newly solved tasks, preserving their evaluated payloads and recorded result columns. +We keep tasks with at least one verifier-successful attempt from the new model. `tasks.parquet` contains only newly kept tasks; tasks already kept in v2 remain separate. ## Evaluation settings -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | On | 2 | 1.0 | 0.95 | 20 | 32,768 | 8,192 | min_p=0; presence_penalty=0; repetition_penalty=1 | -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3.8-27B | 32768 | 8192 | Top-level `max_tokens` only in all 2,118 saved trial configs; not forwarded as `llm_call_kwargs.max_tokens`; actual fallback cap unknown | on | {"min_p": 0.0, "presence_penalty": 0.0, "repetition_penalty": 1.0, "temperature": 1.0, "top_k": 20, "top_p": 0.95} | +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. +**Time limits:** agent 900 seconds; verifier 240 seconds (sampled configurations; limits vary by task where multiple values are shown). -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Thus this run did not request the intended explicit cap. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 enforced tokens, or as unlimited. The current forwarding fix does not retroactively change these results. +**Output-limit evidence:** All 2,118 saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. -All 2,118 saved trial configurations have the same model, sampling and limits. Agent and verifier time limits are task-specific and unchanged from v1. Settings and evidence: [audit](metadata/evaluation-settings-audit.json). +Network isolation was requested, but the run fell back to host networking when network namespace creation failed; agents could have reached the network. -## Evaluation results +## Results -| Model | Tasks evaluated | Attempts | Tasks solved at least once | Pass@1 | Pass@2 | -| --- | ---: | ---: | ---: | ---: | ---: | -| Qwen/Qwen3.8-27B (thinking) | 1,059 | 2,118 | 230 | 18.04% | 21.72% | +The tables and plot cover all evaluated tasks for each model, before selecting tasks for publication. Only the new v3 evaluation is included. -These results cover only the new evaluation of the tasks v2 left unsolved; earlier results are in [v2](../termigen_v2/README.md). +131 additional v1 tasks were outside the earlier 3,336-task evaluation and were not evaluated in this v3 run. -### Successful attempts in the new evaluation - -| Successful attempts | Tasks | +| Successful attempts | Qwen/Qwen3.8-27B (1,059 tasks) | | --- | ---: | | 0/2 | 829 | | 1/2 | 78 | | 2/2 | 152 | -## Select tasks by successful-attempt count +| Model | Evaluated tasks | Pass@1 | Pass@2 | +| --- | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | 1,059 | 18.04% | 21.72% | -The Parquet preserves the v2 columns and adds `qwen38_attempts`, `qwen38_successes`, `qwen38_missing_attempts` and `qwen38_saved_traces`. Tasks already in v2 are not duplicated in this file. +Pass@k uses the sampling-without-replacement estimator from the recorded per-task success counts. At the full attempt budget, it is the fraction of evaluated tasks with at least one observed success. Recorded terminal errors follow the saved scoring policy; absent attempts are not fabricated and raw missing verifier rewards remain null. -```python -import pyarrow.parquet as pq +![Pass rates across all evaluated tasks](plots/readme-pass-at-k.png) -# Newly solved tasks solved exactly once. -subset = pq.read_table("tasks.parquet", filters=[("qwen38_successes", "=", 1)]) -``` +## Files and usage -[Selection metadata](metadata/selection.json) records the status of all 3,467 v1 tasks, the v2 entry of each, and every new attempt. +`tasks.parquet` contains 230 newly kept tasks only. Tasks already kept in v2 are not duplicated. -## Traces +`traces/` contains attempt records across the evaluated cohort, including unsuccessful attempts and tasks not kept. Rows retain task/model/attempt identifiers, raw results and conversations when available. -`traces/` holds all 2,118 new attempts, successful and failed. Historical attempts remain in v2, including the 64 missing historical conversations identified there. Raw `trajectory` and `result` JSON are included; `retained_v3` identifies the incremental task set. +Success-count columns: `instruct_successes`, `qwen38_successes`. Counts are model-specific. The settings above describe this release's evaluation; historical columns retain their earlier budgets. ```python -import json -import pyarrow.dataset as ds +from huggingface_hub import hf_hub_download +import pyarrow.parquet as pq -traces = ds.dataset("traces", format="parquet", exclude_invalid_files=True) -new = traces.to_table(filter=ds.field("evaluation_version") == "v3") -conversation = json.loads(new["trajectory"][0].as_py()) +path = hf_hub_download( + "FWeindel/validated-tasks", + "termigen_v3/tasks.parquet", + repo_type="dataset", +) +# Select kept tasks with exactly one observed success. +tasks = pq.read_table(path, filters=[("qwen38_successes", "=", 1)]) ``` -All 2,118 new attempts have saved conversations. 382 have verifier reward 1, 1,703 verifier reward 0, and 33 no verifier reward. 451 attempts hit the agent timeout; 35 of them still passed the verifier. The 33 attempts without a verifier reward ended in terminal errors (22 tmux protocol errors, 8 task memory-limit errors, 2 runtime errors, 1 verifier timeout); the run's `terminal-errors-zero-v2` policy counts them as zero while preserving their raw null rewards. Traces are included without quality filtering. - -[Trace statistics](trace_stats/README.md) cover only the new evaluation. See [v2 statistics](../termigen_v2/trace_stats/README.md) for earlier models. - -## Limitations - -- **Network:** isolation was requested, but every job fell back to host networking because the nodes had reached their network-namespace limit. Agents could have reached the network. -- **Repository-leak checks:** the checks added on 2026-10-17 (leftover git history and file timestamps that could reveal the solution) have not been applied to these tasks. - -## Runtime - -- **Versions:** Apptainer **1.5.4**, Python **3.12.14**, Harbor **0.8.1**, vLLM **0.30.0**. -- **Implementation:** [Marin Harbor `7faf878c14b6`](https://github.com/marin-community/harbor/tree/7faf878c14b6d72737579ec936ebf5f74ba3194d) with the recorded runtime patches. -- **Resources:** one H200 and 32 CPUs per job, up to 4 concurrent trials per GPU, 8 sequence slots in the serving process. Task CPU, memory and timeout limits are unchanged. - -`patch.py` rebuilds the incremental task file from the pinned pre-migration release and the published layout manifest. - -## Incremental release - -V3 contains only tasks newly solved after v2. It has no task-ID overlap with [v2](../termigen_v2/README.md). This is a harder subset relative to the earlier evaluation, not a claim of intrinsic difficulty. New-evaluation traces include both solved and unsolved tasks so the reported pass rates remain reproducible. +Combine the 2,277 tasks in `termigen_v2/tasks.parquet` with these 230 newly kept tasks for 2,507 unique tasks. Allow for different columns and retain each source file's image manifest when staging images. ```python -from huggingface_hub import hf_hub_download import pyarrow as pa -import pyarrow.parquet as pq def load(version): - path = hf_hub_download("FWeindel/validated-tasks", - f"termigen_{version}/tasks.parquet", repo_type="dataset") + path = hf_hub_download( + "FWeindel/validated-tasks", + f"termigen_{version}/tasks.parquet", + repo_type="dataset", + ) return pq.read_table(path) v2, v3 = load("v2"), load("v3") @@ -105,4 +82,14 @@ assert set(v2["path"].to_pylist()).isdisjoint(v3["path"].to_pylist()) combined = pa.concat_tables([v2, v3], promote_options="permissive") ``` -The example combines task rows; each source file retains its own image manifest. Use those source manifests when staging images. [Layout and provenance](metadata/layout.json) records the pinned baseline and task counts. Earlier cumulative v3 releases remain accessible through their HF commit history. +Repository-history and timestamp leak checks remain incomplete for these tasks. Teacher success does not establish completion of all current datasource-validation checks. + +## Trace statistics + +Of 2,118 recorded attempts, 2,118 have saved conversations and 0 do not. 33 attempts have no raw verifier reward. Missing measurements remain missing. + +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. + +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. Historical attempts remain in v2. + +`trace_stats/README.md` contains detailed turns, token usage, context, runtime, outcome and command distributions. Token totals sum calls across an attempt and are not the context window. diff --git a/termigen_v3/plots/readme-pass-at-k.png b/termigen_v3/plots/readme-pass-at-k.png new file mode 100644 index 0000000000000000000000000000000000000000..a1dda4cfd31de2411cae6f45a5e745f568369cb1 --- /dev/null +++ b/termigen_v3/plots/readme-pass-at-k.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d369c6e35836e8253bb0bb60ccc42243b04f8a379016fd1025e8bb46b482bc30 +size 36161 diff --git a/termigen_v3/trace_stats/README.md b/termigen_v3/trace_stats/README.md index acf33aaf5ece5b569be538f83ed080d7aac6479e..6ed817fb324b0e37aa4a7eaf10648f377c561a4d 100644 --- a/termigen_v3/trace_stats/README.md +++ b/termigen_v3/trace_stats/README.md @@ -1,51 +1,53 @@ -# Trajectory statistics +# TermiGen v3 trace statistics -2,118 attempts across 1,059 tasks of TermiGen v3, Qwen/Qwen3.8-27B evaluation, kept and excluded, each counted once. Task success counts are in the [dataset overview](../README.md). +## Coverage -| Metric | Qwen/Qwen3.8-27B | -| --- | ---: | -| Recorded attempts | 2,118 | -| Median agent turns | 8 | -| Median input tokens per attempt | 55,973 | -| Median output tokens per attempt | 16,981 | -| Median agent runtime (seconds) | 335 | -| Median verifier runtime (seconds) | 10 | +Of 2,118 recorded attempts, 2,118 have saved conversations and 0 do not. 33 attempts have no raw verifier reward. Missing measurements remain missing. -Medians use recorded measurements only. This page covers the new v3 evaluation; [v2 statistics](../../termigen_v2/trace_stats/README.md) cover the historical traces. +Agent timeouts and verifier outcomes are separate: an attempt can time out and still receive a successful verifier grade. -## Plots +The export includes successful and unsuccessful attempts on both kept and unselected tasks. Conversation availability is not a training-quality acceptance claim; no additional conversation quality filter is applied to this export. Historical attempts remain in v2. -![Trace measurements](plots/all-models-trace-distributions.png) +Token totals sum calls across an attempt and are not the context window. -![Trace outcomes](plots/all-models-trace-outcomes.png) +Measurements cover 1,059 tasks, including kept and unselected tasks, with each attempt counted once. -![Most frequent commands](plots/all-models-trace-commands.png) +## Evaluation settings -Commands are counted as submitted, not as successful executions. +| Model | Agent | Thinking | Attempts/task | Temperature | Top-p | Top-k | Context tokens | Intended output tokens/call | Other sampling | +| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | +| Qwen/Qwen3.8-27B | Terminus-2 | On | 2 | 1.0 | 0.95 | 20 | 32,768 | 8,192 | min_p=0; presence_penalty=0; repetition_penalty=1 | -Per-model plots: [Qwen/Qwen3.8-27B commands](plots/qwen38-27b-trace-commands.png), [Qwen/Qwen3.8-27B distributions](plots/qwen38-27b-trace-distributions.png), [Qwen/Qwen3.8-27B outcomes](plots/qwen38-27b-trace-outcomes.png). +Context includes prompt and response for one model call. The output budget includes thinking tokens; neither is a whole-attempt token total. -## Data +**Time limits:** agent 900 seconds; verifier 240 seconds (sampled configurations; limits vary by task where multiple values are shown). -- [Summary](data/summary.json): overall and per-model statistics. -- [Per-task metrics](data/per_task.parquet) and [per-attempt metrics](data/per_attempt.parquet), with nested metrics stored as JSON. -- [Provenance](data/provenance.json): source files and deduplication. +**Output-limit evidence:** All 2,118 saved configurations use a top-level output limit that the affected Terminus path did not forward as an explicit request cap. The actual fallback cap is unknown. The later forwarding fix does not change these recorded results. -Attempts without a saved conversation: 0; without a verifier reward: 33. Their missing measurements stay empty. +Network isolation was requested, but the run fell back to host networking when network namespace creation failed; agents could have reached the network. +## Measurements -## Evaluation settings +### Qwen/Qwen3.8-27B -Context is the total prompt plus response window for one model call. The response budget includes reasoning tokens and is not a whole-attempt token budget. +| Measurement | Attempts measured | Median | 90th percentile | +| --- | ---: | ---: | ---: | +| Agent turns | 2,118 | 8 | 16 | +| Input tokens per attempt | 2,118 | 55,973 | 175,690 | +| Output tokens per attempt | 2,118 | 16,981 | 45,307 | +| Largest recorded step, prompt plus output tokens | 2,117 | 15,662 | 29,659 | +| Agent runtime (seconds) | 2,118 | 335.39 | 900.01 | +| Verifier runtime (seconds) | 2,088 | 9.8 | 11.91 | +| Whole-attempt runtime (seconds) | 2,118 | 352.61 | 923.26 | -| Model | Context tokens (configured) | Intended response tokens per call | Output-limit evidence | Thinking | Sampling | -| --- | ---: | ---: | --- | --- | --- | -| Qwen/Qwen3.8-27B | 32768 | 8192 | Top-level `max_tokens` only in all 2,118 saved trial configs; not forwarded as `llm_call_kwargs.max_tokens`; actual fallback cap unknown | on | {"min_p": 0.0, "presence_penalty": 0.0, "repetition_penalty": 1.0, "temperature": 1.0, "top_k": 20, "top_p": 0.95} | +Statistics use available measurements only. Input/output totals sum calls and can include repeated context. Recorded ATIF steps can aggregate multiple requests, so their largest prompt-plus-output count does not establish a per-call context limit. Missing summarization counts are not zero. Commands are counted as submitted, not as successful executions. -Configured limits are not measurements of token usage. A missing explicit response cap does not mean unlimited generation: the serving context, backend defaults, stopping conditions and task timeouts still apply. +Machine-readable measurements remain in `data/summary.json`, `data/per_task.parquet` and `data/per_attempt.parquet`. -**Historical output-limit bug:** the intended per-response limit was stored as top-level `max_tokens`, but the affected Terminus path did not forward it as `llm_call_kwargs.max_tokens`. Generation instead depended on the backend's fallback allowance, available context, stopping conditions and agent timeout. The exact fallback limit is not established by the saved configurations; it must not be reported as 8,192 enforced tokens, or as unlimited. The current forwarding fix does not retroactively change these results. +## Plots -Agent and verifier time limits are task-specific; see [settings and evidence](../metadata/evaluation-settings-audit.json). +![Trace measurements](plots/all-models-trace-distributions.png) -This folder reports the new evaluation only. Task membership is incremental: tasks already in v2 are not duplicated in v3. Historical conversations remain in v2. See the [release overview](../README.md) for counts and combination instructions. +![Trace outcomes](plots/all-models-trace-outcomes.png) + +![Most frequent commands](plots/all-models-trace-commands.png)