review-artifacts/nsdl-7b-rl-locus
014
NSDL 7B (SFT + RL), LOCUS: anonymized checkpoint for double-anonymous review
Row "Table 2: NSDL 7B (SFT + RL); Fig. 3" of the submission. Weights are stored in bfloat16, the dtype the evaluation loader uses. Load and run with the code in the anonymized repository linked from the paper (nsdl/ + scripts/eval_lnsdl_agentic_viki_l2.py); evaluation flags are in that repository's README. Step-30 checkpoint of the RL run initialised from the NSDL 7B (SFT) LOCUS checkpoint (scripts/rlfinalsetting.sh). Provided for peer review only; do not redistribute during review. Author, affiliation and citation information is intentionally omitted.
