Skip to content

docs: add judge validation and eval case count to the README #26

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ana-lan merged 1 commit into from
Nov 1, 2026
Merged

docs: add judge validation and eval case count to the README #26

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 5 additions & 1 deletion README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -81,7 +81,11 @@ SupportDesk (`examples/supportdesk`) is a synthetic support ticket app with 8 LL
| Models resolved | 6 | 8 |
| Prompts resolved | 5 | 8 |

After evals on local Qwen 2.5 models (7b baseline; 3b, 1.5b, 0.5b candidates), **3 of 8 call sites can downgrade**. Projected monthly cost goes from **$3,137.26 to $2,572.61, saving $564.65 (18%)**, while the mean pass rate moves from 68.3% to 68.9%. Full report: [`examples/supportdesk/downshift.report.md`](https://github.com/ana-lan/downshift/blob/main/examples/supportdesk/downshift.report.md). Bob then applied the decisions and opened the PR itself.
After 182 eval cases across the 8 call sites on local Qwen 2.5 models (7b baseline; 3b, 1.5b, 0.5b candidates), **3 of 8 call sites can downgrade**. Projected monthly cost goes from **$3,137.26 to $2,572.61, saving $564.65 (18%)**, while the mean pass rate moves from 68.3% to 68.9%. Full report: [`examples/supportdesk/downshift.report.md`](https://github.com/ana-lan/downshift/blob/main/examples/supportdesk/downshift.report.md). Bob then applied the decisions and opened the PR itself.

### Judge validation

Two call sites write free text (`draft_reply`, `summarize_for_agent`), so an LLM judge grades them. When the 7b baseline graded its own outputs, `draft_reply` passed 22/22 (100%) and `summarize_for_agent` passed 20/22 (91%). Re-grading the same saved outputs with `openai/gpt-oss-120b` via `downshift rescore` dropped them to 3/22 (14%) and 13/22 (59%). All reported results use the independent judge. Both runs are in the [raw results](https://github.com/ana-lan/downshift/blob/main/examples/supportdesk/results/supportdesk.agent_assist__draft_reply/qwen2.5-7b.jsonl) (rows 1-22 self-graded, 23-44 rescored).

## Case studies on real repos

Expand Down
Loading