jevify_config.json

{
 "jevify_version": 1,
 "tier": 2,
 "method": "readout-lora",
 "backbone": "HuggingFaceTB/SmolLM3-3B-checkpoints",
 "backbone_revision": "cfb32d505f5025ec9be4e704f70cfbf5bdf8da94",
 "base_model": "HuggingFaceTB/SmolLM3-3B-checkpoints",
 "chat": true,
 "recipe": {
  "mode": "index",
  "permutations": {
   "choice": 2,
   "score": 1,
   "noul": 1
  },
  "prior_weight": {
   "choice": 0.0,
   "score": 0.0,
   "noul": 0.0
  },
  "temperature": {
   "choice": 1.2192,
   "score": 1.4924,
   "noul": 1.8579
  },
  "bias": {
   "noul": -0.2
  },
  "state_last": false,
  "prompt": "jevify"
 },
 "training": {
  "objective": "proper scoring rule on the readout",
  "arm": "sup",
  "coherence_beta": 0.0,
  "seed": 0,
  "lr": 3e-05,
  "epochs": 2,
  "best_epoch": 1,
  "best_val_loss": 0.6561664592338493,
  "records_per_source_cap": 400,
  "n_train_families": 5885,
  "full_fine_tune": false,
  "precision": "bf16",
  "trained_on": "Praveenrajus/jev-bench train splits of the 16 sources that are not held out",
  "heldout_sources": [
   "clinc150",
   "arc_challenge",
   "yelp5",
   "measuring_hate_speech",
   "fever_evidence",
   "strategyqa_grounded"
  ]
 },
 "evaluated_on": "Praveenrajus/jev-bench (test, 22,773 records)",
 "lora": {
  "r": 16,
  "alpha": 32,
  "target_modules": [
   "gate_proj",
   "q_proj",
   "up_proj",
   "v_proj",
   "down_proj",
   "k_proj",
   "o_proj"
  ],
  "trainable": 30228480,
  "merged_at_load": true
 }
}