Md Saiful Islam Sajol
Claude Opus 4.8
Sync to GitHub main fd62f0f: THRIVE v6.1 weights, v6/TriviaQA pipeline, labelling tooling
ac442f8 | # Overnight queue: wait for the crosstask grid (15 fits) to finish, then run | |
| # the v6.1 rebind: full suite retrain on the canonical hallu_v6_1_eval.jsonl | |
| # at the FINAL v6 config (bs2 / ap_pairwise mix 0.75 / ep15). Signals are extracted FRESH | |
| # from the canonical v6.1 file into outputs/v6_1_signals (researcher call | |
| # 2026-09-18: full provenance chain, no reuse seam; +6-8h). | |
| # | |
| # tmux new-session -d -s v61 'bash scripts/queue_v61_rebind.sh 2>&1 | tee outputs/queue_v61.log' | |
| set -u | |
| cd "$(dirname "${BASH_SOURCE[0]}")/.." | |
| echo "[queue] waiting for crosstask grid $(date -Is)" | |
| while true; do | |
| n=$(ls outputs/rt_ctmix_s*/latticespike_*/summary.json 2>/dev/null | wc -l) | |
| active=$(find outputs/rt_ctmix_s4* -name run.log -mmin -30 2>/dev/null | wc -l) | |
| if [[ "$n" -ge 15 ]]; then echo "[queue] grid complete ($n fits) $(date -Is)"; break; fi | |
| if [[ "$active" -eq 0 ]]; then echo "[queue] no grid activity 30min ($n/15) -- proceeding $(date -Is)"; break; fi | |
| sleep 120 | |
| done | |
| sleep 120 # let the grid collector finish printing | |
| echo "[queue] launching v6.1 rebind suite $(date -Is)" | |
| DATA="$HOME/datasets/hallucinations_custom/hallu_v6_1_eval.jsonl" \ | |
| SPLIT="$HOME/datasets/hallucinations_custom/hallu_v6_1_split.json" \ | |
| SIGNALS="outputs/v6_1_signals" \ | |
| ROOT="$PWD/outputs/v6_1_methods" \ | |
| bash scripts/run_v6_all.sh 2>&1 | tee outputs/v6_1_all.log | |
| echo "[queue] rebind finished $(date -Is)" | |