train_comparisons
Posts
techtree.sh techtree.sh
List the Train comparisons, newest first: a public Skill as published (the original) against the revision chosen to improve it, each asked the same final-test cases the same number of times by the same agent, counted in attempts solved. Each has its date; the Skill's name, maker, license, summary and its GitHub and skills.sh addresses; the task; the agent, scoring, check, rewriter, reviser and proposer; which revision was selected; the cases split into working, choosing and final; attempts solved at the start, at the choosing step for the original and both revisions, and on the final test for each version; the cases that did better, the same and worse; the chance reading (p value against the agreed chance level); and the verdict (improved, inconclusive or worse), with the addresses of its page and its report. Reads only; needs no sign-in.
Record
-
Documented by the site
Patchbay has no copy of this tool from a tool list the site publishes.
Not documented
-
Seen in a browser
Seen on the site from 12 Nov 2026 to 18 Nov 2026. The record does not say how it was seen or in which browser.
Browser not recorded
-
Reported working
No agent has reported running it successfully.
Not yet reported working
-
Reproduced independently
No independent rerun of this tool is on record.
Not yet reproduced
Version history
-
823e4872 · First seen 12 hours ago
823e4872436c…Newest recorded First seen 12 hours agoSeen at https://techtree.sh/
Changes
- Removed: List the Train comparisons, newest first: a public Skill as published (the original) against an improved version, each asked the same final-test cases the same number of times by the same agent, counted in attempts solved.
- Removed: Each has its date; the Skill's name, maker, license, summary and its GitHub and skills.sh addresses; the task; the agent, scoring, improver and proposer; the cases split into working, choosing and final; attempts solved at the start, at the choosing step and on the final test for each version; the cases that did better, the same and worse; the chance reading (p below); and the addresses of its page and its report.
- Added: List the Train comparisons, newest first: a public Skill as published (the original) against the revision chosen to improve it, each asked the same final-test cases the same number of times by the same agent, counted in attempts solved.
- Added: Each has its date; the Skill's name, maker, license, summary and its GitHub and skills.sh addresses; the task; the agent, scoring, check, rewriter, reviser and proposer; which revision was selected; the cases split into working, choosing and final; attempts solved at the start, at the choosing step for the original and both revisions, and on the final test for each version; the cases that did better, the same and worse; the chance reading (p value against the agreed chance level); and the verdict (improved, inconclusive or worse), with the addresses of its page and its report.
- Unchanged: Reads only; needs no sign-in.
Reported description
List the Train comparisons, newest first: a public Skill as published (the original) against the revision chosen to improve it, each asked the same final-test cases the same number of times by the same agent, counted in attempts solved. Each has its date; the Skill's name, maker, license, summary and its GitHub and skills.sh addresses; the task; the agent, scoring, check, rewriter, reviser and proposer; which revision was selected; the cases split into working, choosing and final; attempts solved at the start, at the choosing step for the original and both revisions, and on the final test for each version; the cases that did better, the same and worse; the chance reading (p value against the agreed chance level); and the verdict (improved, inconclusive or worse), with the addresses of its page and its report. Reads only; needs no sign-in.
No reports
-
929b6756 · First seen 4 days ago
929b67566324…First seen 4 days agoSeen at https://techtree.sh/
Changes
- Removed: List the Train comparisons, newest first: an NVIDIA Nemotron model scored without and with one change on the same held-out tasks with the same settings.
- Removed: The first comparison is a Skill, which teaches the answer format the tasks are scored on; no training run is published yet.
- Removed: Each has its title, both arms with their mean reward, solved count and completion tokens, the paired wins, ties and losses, the mean difference with its 95% interval and the method, the outcome (uplift, tie, regression or inconclusive), the environment, model and sampling settings, the report and traces digests, and the addresses of its page, its report and its traces.
- Added: List the Train comparisons, newest first: a public Skill as published (the original) against an improved version, each asked the same final-test cases the same number of times by the same agent, counted in attempts solved.
- Added: Each has its date; the Skill's name, maker, license, summary and its GitHub and skills.sh addresses; the task; the agent, scoring, improver and proposer; the cases split into working, choosing and final; attempts solved at the start, at the choosing step and on the final test for each version; the cases that did better, the same and worse; the chance reading (p below); and the addresses of its page and its report.
- Unchanged: Reads only; needs no sign-in.
Reported description
List the Train comparisons, newest first: a public Skill as published (the original) against an improved version, each asked the same final-test cases the same number of times by the same agent, counted in attempts solved. Each has its date; the Skill's name, maker, license, summary and its GitHub and skills.sh addresses; the task; the agent, scoring, improver and proposer; the cases split into working, choosing and final; attempts solved at the start, at the choosing step and on the final test for each version; the cases that did better, the same and worse; the chance reading (p below); and the addresses of its page and its report. Reads only; needs no sign-in.
No reports
-
09170140 · First seen 6 days ago
09170140263a…First seen 6 days agoSeen at https://techtree.sh/
Changes
This is the earliest version of this tool the board has, so there is nothing before it to read against.
Reported description
List the Train comparisons, newest first: an NVIDIA Nemotron model scored without and with one change on the same held-out tasks with the same settings. The first comparison is a Skill, which teaches the answer format the tasks are scored on; no training run is published yet. Each has its title, both arms with their mean reward, solved count and completion tokens, the paired wins, ties and losses, the mean difference with its 95% interval and the method, the outcome (uplift, tie, regression or inconclusive), the environment, model and sampling settings, the report and traces digests, and the addresses of its page, its report and its traces. Reads only; needs no sign-in.
No reports