CCT Summarizer V0.2 · LoRA Eval loading

Qwen3-30B-A3B base vs our LoRA fine-tunes on turns never seen in training, every summarizer run side by side, training and spend. Live from the run data.

base / other models LoRA v2 (ours) reference (Haiku 4.5)

Human-review audit V0.1 archive

—

touched-task F1 · LoRA

—

valid output · LoRA

—

held-out val rows (turns 1203-1335) · bf16 H100

—

paid API spend (TF + OpenRouter ledgers)

Headline · base vs LoRA on the held-out val — loading

Rollout leaderboard — each model labels the day on its own, from turn 1

Coverage = turns labeled of the window. turns/h excludes pauses between calls. USD is each run's own call log (Haiku: list-price equivalent on the subscription, not paid). Run names open in the timeline.

swarm-board V0.2static demo

CCT Summarizer V0.2 · LoRA Eval loading

Headline · base vs LoRA on the held-out val — loading

Rollout leaderboard — each model labels the day on its own, from turn 1

Training — Token Factory LoRA jobs

Spend — paid runs, from the spend ledgers