Rebuild trace-label join; add Qwen/CodeLlama traces (1642 rows)
Browse filesmast_annotation was previously indexed by row position rather than joined to the trace, so all configurations shared one 206-long label sequence. Labels are rebuilt from the raw judge outputs and joined per trace. Adds 400 open-weight rows (Qwen, CodeLlama on ProgramDev-v2). Adds a uid primary key.
- MAD_full_dataset.json +2 -2
- README.md +1 -5
MAD_full_dataset.json
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d636ac63dfc1c6af2d312e862f4b7d383b62d3898d431ccd0e7a79d21d85406f
|
| 3 |
+
size 199574367
|
README.md
CHANGED
|
@@ -40,11 +40,10 @@ Code: https://github.com/multi-agent-systems-failure-taxonomy/MAST
|
|
| 40 |
|
| 41 |
| field | type | |
|
| 42 |
|---|---|---|
|
| 43 |
-
| `uid` | string | primary key, `"<config>#<index>"` |
|
| 44 |
| `mas_name` | string | AG2, MetaGPT, ChatDev, Magentic, AppWorld, HyperAgent, OpenManus |
|
| 45 |
| `llm_name` | string | GPT-4o, Claude, GPT-4o-mini, Qwen, CodeLlama |
|
| 46 |
| `benchmark_name` | string | ProgramDev, ProgramDev-v2, GSM, Olympiad, GAIA, MMLU, Test-C, SWE-Bench-Lite |
|
| 47 |
-
| `trace_id` | int | index within its config
|
| 48 |
| `trace` | dict | `{key, index, trajectory}` |
|
| 49 |
| `mast_annotation` | dict | 14 codes → `1`, `0`, or `null` where the annotation is unavailable |
|
| 50 |
|
|
@@ -64,9 +63,6 @@ Code: https://github.com/multi-agent-systems-failure-taxonomy/MAST
|
|
| 64 |
|
| 65 |
Annotations are produced by an LLM judge, not by human labelling.
|
| 66 |
|
| 67 |
-
Some trajectories appear more than once under different configurations. Group on
|
| 68 |
-
`trace.trajectory` for 1537 distinct traces.
|
| 69 |
-
|
| 70 |
In `MAD_human_labelled_dataset.json`, each round uses a different revision of the taxonomy
|
| 71 |
(18, 17, 17 and 14 modes) and the codes are not comparable across rounds.
|
| 72 |
|
|
|
|
| 40 |
|
| 41 |
| field | type | |
|
| 42 |
|---|---|---|
|
|
|
|
| 43 |
| `mas_name` | string | AG2, MetaGPT, ChatDev, Magentic, AppWorld, HyperAgent, OpenManus |
|
| 44 |
| `llm_name` | string | GPT-4o, Claude, GPT-4o-mini, Qwen, CodeLlama |
|
| 45 |
| `benchmark_name` | string | ProgramDev, ProgramDev-v2, GSM, Olympiad, GAIA, MMLU, Test-C, SWE-Bench-Lite |
|
| 46 |
+
| `trace_id` | int | index within its config |
|
| 47 |
| `trace` | dict | `{key, index, trajectory}` |
|
| 48 |
| `mast_annotation` | dict | 14 codes → `1`, `0`, or `null` where the annotation is unavailable |
|
| 49 |
|
|
|
|
| 63 |
|
| 64 |
Annotations are produced by an LLM judge, not by human labelling.
|
| 65 |
|
|
|
|
|
|
|
|
|
|
| 66 |
In `MAD_human_labelled_dataset.json`, each round uses a different revision of the taxonomy
|
| 67 |
(18, 17, 17 and 14 modes) and the codes are not comparable across rounds.
|
| 68 |
|