FORWARD PASS 1 of 5:
{
 "ok": true,
 "type": "generate",
 "model": "original",
 "generations": 2048,
 "graded": 2048,
 "accuracy": 0.16455078125,
 "accuracy_by_source": {
  "gsm8k": 0.078125,
  "math": 0.193359375
 },
 "mean_tokens": 1197.73193359375,
 "cut_share": 0.06591796875,
 "failures": {
  "cut_at_max_tokens": 129,
  "no_answer_finished": 5,
  "wrong_answer": 1577
 },
 "sampling": {
  "temperature": 1.0,
  "top_p": 0.95,
  "top_k": 64,
  "max_tokens": 8192,
  "thinking": false
 },
 "seed_policy": "SHA256(exp8-v1:item_identity:sample_index), first 32 bits little-endian, masked to 31 bits; original mapping unpublished",
 "runtime": {
  "gpu_memory_utilization": 0.35,
  "max_num_seqs": 32,
  "max_model_len": 10240,
  "enforce_eager": true,
  "prefix_caching": false
 },
 "grading": {
  "mathqa_sha256": "f0d576a1adc625dede7bff749b3bf8c9a5a8f6bae43d51e2d0ee36d694e89843",
  "evaluate_sha256": "c93f477c02b3ae0d6b37abe3afad8bb7353e273322ff963eec7ff2270012035d",
  "math_verify_version": "0.9.0",
  "gemma_lenient_functions": "unchanged AST bodies from original evaluate.py",
  "loop_signal_deviation": "Missing mcqa repetitive_end helper reconstructed from the exact 200-word/4-gram/three-repeat/half-position rule documented in mathqa.py."
 },
 "s": 20667.92,
 "file": "/work/gen_2.jsonl",
 "activations": {
  "dir": "/work/gen_2_acts/",
  "files": {
   "hidden_last_prompt.npy": [
    2048,
    19,
    640
   ],
   "hidden_mean_answer.npy": [
    2048,
    19,
    640
   ],
   "mlp_last_prompt.npy": [
    2048,
    18,
    2048
   ],
   "mlp_mean_answer.npy": [
    2048,
    18,
    2048
   ]
  }
 }
}