Solshine/gemma-4-e2b-nla-L23-av-priordev-wd3-20k
Publish model card + evaluation results (eval_results.json)
Publish model card + evaluation results (README.md)
Upload README.md with huggingface_hub
Upload eval_data/balanced_eval_txt.parquet with huggingface_hub
Upload eval_results/traj_acc_act_swap_n580.json with huggingface_hub
Upload eval_results/f124_power_eval_n580.json with huggingface_hub
add powered_eval_heldout.parquet (n=580, L23 activations, 3 domains) for §F124 power run
trajectory CSV after step 20000
eval step 20000 tfidf=0.135
trajectory CSV after step 19000
eval step 19000 tfidf=0.115
trajectory CSV after step 18000
eval step 18000 tfidf=0.096
trajectory CSV after step 17000
eval step 17000 tfidf=0.173
trajectory CSV after step 16000
eval step 16000 tfidf=0.173
trajectory CSV after step 15000
eval step 15000 tfidf=0.096
trajectory CSV after step 14000
eval step 14000 tfidf=0.115
trajectory CSV after step 13000
eval step 13000 tfidf=0.077
trajectory CSV after step 12000
eval step 12000 tfidf=0.173
trajectory CSV after step 11000
eval step 11000 tfidf=0.173
trajectory CSV after step 10000
eval step 10000 tfidf=0.250
trajectory CSV after step 9000
eval step 9000 tfidf=0.173
trajectory CSV after step 8000
eval step 8000 tfidf=0.135
trajectory CSV after step 7000
eval step 7000 tfidf=0.212
trajectory CSV after step 6000
eval step 6000 tfidf=0.096
trajectory CSV after step 5000
eval step 5000 tfidf=0.173
trajectory CSV after step 4000
eval step 4000 tfidf=0.250
trajectory CSV after step 3000
eval step 3000 tfidf=0.192
trajectory CSV after step 2000
eval step 2000 tfidf=0.135
trajectory CSV after step 1000
eval step 1000 tfidf=0.135
test write access
NLA prior-dev wd=0.3 20000k-step grokking run
trainloss up to step 20000