Datasets:
Update README.md
Browse files
README.md
CHANGED
|
@@ -73,21 +73,21 @@ Trace stats group each attempt by how the agent stopped:
|
|
| 73 |
| Outcome | Meaning |
|
| 74 |
| --- | --- |
|
| 75 |
| Task complete | The agent said it was done and confirmed that it wanted to finish. The verifier still checks whether the solution is correct. |
|
| 76 |
-
| tmux session closed | The agent's terminal closed before it confirmed
|
| 77 |
| Task timeout | The agent ran out of time. |
|
| 78 |
-
| Repeated output truncation | Three responses in a row were too long and could not be used. |
|
| 79 |
| Context recovery failure | The harness could not make enough room in the conversation for the model to continue. |
|
| 80 |
| Other limits | The agent hit another limit, such as the number of turns, response length, context size or task memory. |
|
| 81 |
| Error | The run stopped because of an error that does not fit a more specific outcome above. |
|
| 82 |
| Not recorded | The saved result does not say why the agent stopped. |
|
| 83 |
|
| 84 |
-
These labels come from the harness's recorded stop reason and error type, not an LLM judge
|
| 85 |
|
| 86 |
Recorded errors are also grouped by where they happened:
|
| 87 |
|
| 88 |
| Error category | Meaning |
|
| 89 |
| --- | --- |
|
| 90 |
-
| Environment / terminal | Something went wrong with the container or terminal. The agent may have caused it
|
| 91 |
| Verifier | The checker could not run or finish normally. A solution simply failing its tests is not a verifier error. |
|
| 92 |
| Model server | A request to the model failed, for example because the server was unavailable or the request timed out. |
|
| 93 |
| Other | An error was recorded, but its type does not fit the categories above. |
|
|
|
|
| 73 |
| Outcome | Meaning |
|
| 74 |
| --- | --- |
|
| 75 |
| Task complete | The agent said it was done and confirmed that it wanted to finish. The verifier still checks whether the solution is correct. |
|
| 76 |
+
| tmux session closed | The agent's terminal closed before it confirmed completionk, for example, if agent writes an `exit` command. |
|
| 77 |
| Task timeout | The agent ran out of time. |
|
| 78 |
+
| Repeated output truncation | Three responses in a row were too long (current default is 8,192) and could not be used. |
|
| 79 |
| Context recovery failure | The harness could not make enough room in the conversation for the model to continue. |
|
| 80 |
| Other limits | The agent hit another limit, such as the number of turns, response length, context size or task memory. |
|
| 81 |
| Error | The run stopped because of an error that does not fit a more specific outcome above. |
|
| 82 |
| Not recorded | The saved result does not say why the agent stopped. |
|
| 83 |
|
| 84 |
+
These labels come from the harness's recorded stop reason and error type, not an LLM judge and describe how the run ended.
|
| 85 |
|
| 86 |
Recorded errors are also grouped by where they happened:
|
| 87 |
|
| 88 |
| Error category | Meaning |
|
| 89 |
| --- | --- |
|
| 90 |
+
| Environment / terminal | Something went wrong with the container or terminal. The agent may have caused it, for example, by killing its own terminal. |
|
| 91 |
| Verifier | The checker could not run or finish normally. A solution simply failing its tests is not a verifier error. |
|
| 92 |
| Model server | A request to the model failed, for example because the server was unavailable or the request timed out. |
|
| 93 |
| Other | An error was recorded, but its type does not fit the categories above. |
|