Dataset Viewer
Auto-converted to Parquet Duplicate
type
stringclasses
2 values
harness
stringclasses
1 value
id
stringclasses
1 value
name
stringclasses
1 value
message
dict
session
trackio
agent_run
Agent repro run 2026-08-03 03:37 UTC
null
message
null
null
null
{ "role": "system", "content": "Event", "timestamp": 1785728255596 }
message
null
null
null
{ "role": "assistant", "content": "[Claim 0 — PI controller learns dynamic Top-p threshold]\n# Claim 0 — PI controller learns dynamic Top-p threshold\n\nEvidence.\n- Numbers from job `6a5ee27ad09dc1f57c6be6f6`:\n - Sparsity low: `0.5`\n - Sparsity high: `0.9`\n- Interpretation: simulated control trace shows thres...
message
null
null
null
{ "role": "assistant", "content": "[Claim 1 — Fixed Top-p matches Top-k MoE FLOPs]\n# Claim 1 — Fixed Top-p matches Top-k MoE FLOPs\n\nEvidence.\n- Numbers from job `6a5ee27ad09dc1f57c6be6f6`:\n - Fixed Top-p approximate active experts: `8`\n - Top-k baseline: `12`\n- A real expert-routing benchmark would be requ...
message
null
null
null
{ "role": "assistant", "content": "[Claim 2 — Strong scaling across expert granularity / capacity / model / dataset size]\n# Claim 2 — Strong scaling across expert granularity / capacity / model / dataset size\n\nEvidence.\n- Only a small control-theoretic proxy was run; scaling sweeps are not present.\n- Numbers f...
message
null
null
null
{ "role": "assistant", "content": "[hEvt3HWxD3 — DTop-p / dynamic MoE routing logbook]\n# hEvt3HWxD3 — DTop-p / dynamic MoE routing logbook\n\nScope: toy CPU proxy for PI-controller threshold learning.\n\nArtifact:\n- Job `6a5ee27ad09dc1f57c6be6f6` (COMPLETED): controller sparsity trace and fixed Top-p equivalent a...
message
null
null
null
{ "role": "system", "content": "Event", "timestamp": 1785728255596 }

Agent traces

Agent sessions published from a Trackio Logbook.

Downloads last month
7