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 03:38 UTC | null |
message | null | null | null | {
"role": "system",
"content": "Event",
"timestamp": 1785728326158
} |
message | null | null | null | {
"role": "assistant",
"content": "[Claim 1: ConceptMoE improves pretraining and task performance]\n# Claim 1: ConceptMoE improves pretraining and task performance\n\n**Claim:** ConceptMoE achieves +0.9 points on language pretraining, +2.3 on long context, and +0.6 on multimodal tasks.\n\n## Real HF job evidence\n\... |
message | null | null | null | {
"role": "assistant",
"content": "[Claim 1: Attention compute and KV cache savings at R=2]\n# Claim 1: Attention compute and KV cache savings at R=2\n\n**Claim:** At compression ratio R=2, ConceptMoE reduces attention compute by R²× and KV cache by R×, delivering prefill speedups up to 175% and decoding speedups u... |
message | null | null | null | {
"role": "assistant",
"content": "[Claim 2: Continual training conversion with layer looping]\n# Claim 2: Continual training conversion with layer looping\n\n**Claim:** Continual training conversion with layer looping gains additional +5.5 points.\n\n## Real HF job evidence\n\nJob: `Edd16/6a5f050dd09dc1f57c6beadd`... |
message | null | null | null | {
"role": "assistant",
"content": "[Reproducing ConceptMoE / yaUJPj48wa (ICML 2026)]\n# Reproducing ConceptMoE / yaUJPj48wa (ICML 2026)\n\nLocal Trackio logbook reproducing three claims from paper `yaUJPj48wa` with a real HuggingFace Jobs run on `cpu-basic`.\n\n## HF Job evidence\n\n| Field | Value |\n|---|---|\n| ... |
message | null | null | null | {
"role": "system",
"content": "Event",
"timestamp": 1785728326158
} |
- Downloads last month
- 5