{
  "system": "Gemma 4 E4B instruction-tuned model (text-only) in thinking mode with a LoRA adapter trained by STaR on its own thoughts; each question is answered in all 4 choice orders with a thought of up to 512 tokens, easy questions once more with a thought of up to 2,048 tokens as a fifth vote, medium questions also in the 4 reversed choice orders, hard questions once more with a thought of up to 2,048 tokens that replaces their vote, and the answer is the highest average probability among the option numbers after the thought",
  "components": [
    {
      "role": "generator",
      "model": "google/gemma-4-E4B-it",
      "url": "https://huggingface.co/google/gemma-4-E4B-it",
      "license": "Apache-2.0",
      "released": "2026-03-02",
      "loaded_parts": "language model only (vision and audio encoders are not loaded)",
      "variant": "instruction-tuned (-it), thinking mode on with a greedy thought of up to 512 tokens on every question in 4 choice orders (predict.py --rotations 0 2 and --rotations 1 3) and on medium questions in the 4 reversed orders (predict.py --level medium --reverse), and of up to 2,048 tokens on easy and hard questions in the given order (predict.py --level easy|hard --think 2048)",
      "parameters_loaded": 7463013376,
      "parameters_full_checkpoint": 7996156490,
      "precision": "float16 (run.sh sets it on every GPU; predict.py alone picks bfloat16 on GPUs with compute capability 8 or more)",
      "fine_tuned": false
    },
    {
      "role": "adapter",
      "model": "LoRA adapter made by train_lora.py",
      "method": "LoRA r=16, alpha 32, dropout 0.05 on every linear layer of the 42 decoder layers (342 modules), merged into the base weights at load",
      "training": "STaR round 3: the model without any adapter thinks about each of 8,894 verified training questions; star.py --no_hint_retry keeps the thoughts that finish within 512 tokens and give the verified answer on the first try (4,073); a new adapter is trained from the base model on them, 2 epochs. Nothing from the round 1 or round 2 adapters is used",
      "parameters": 38666240,
      "optional": true
    }
  ],
  "total_inference_parameters": {
    "without_adapter": 7463013376,
    "with_adapter": 7501679616
  },
  "parameter_limit": 8000000000,
  "retrieval": "none",
  "internet_at_inference": false,
  "training_data": "data/train/verified-mcq.jsonl: 10,000 of the 18,398 Sinhala multiple-choice questions in data/train/mcq.jsonl (generated from Sri Lankan school textbooks, NIE teacher guides and resource books, A/L learning outcome notes and Wikibooks), with answer keys checked by a stronger model. review_overlap.py removes the 1,106 that match a dev or test question by wording or meaning, and the adapter trains only on the first-try thoughts of the model without an adapter that star.py keeps for the other 8,894. The dev and test sets are not used for training.",
  "closed_api_usage": "Anthropic claude-haiku-5-5 generated the training questions in data/train/mcq.jsonl, and claude-sonnet-5-5 and claude-opus-5-5 checked the answer keys in data/train/verified-mcq.jsonl. They are used for training data only. No closed API is used at inference."
}