{
  "_doc": "EXPERIMENT: Options 1+3 = richer InfoNCE target (full y, max_len=128) + MLP projector (d->4d->d with GELU). Tests whether stronger supervision + more expressive compression fixes the absolute-accuracy ceiling we hit at 7B. Block_z_to_x=False because pilot 7B AR Δ_random=13pp already shows the architecture is content-load-bearing in AR (TF just understates it); the real problem is absolute accuracy (13% on GSM8K). 2500 K=8 steps then 1000 K=16 steps (3500 total) to match pilot 7B's compute, but with the two new mechanisms. Hypothesis: richer signal + MLP pushes absolute acc materially above pilot 7B's 13%.",

  "base_model": "Qwen/Qwen2.5-Math-7B-Instruct",
  "use_lora": true,
  "lora_r": 16,
  "lora_alpha": 32,
  "lora_dropout": 0.05,
  "lora_target_modules": ["q_proj", "k_proj", "v_proj", "o_proj"],
  "dtype": "bfloat16",
  "attn_impl": "eager",
  "gradient_checkpointing": false,

  "K_latents": 8,
  "K_curriculum": [[0, 8], [2500, 16]],
  "block_y_to_x": true,
  "block_z_to_x": false,
  "proj_init_scale": 0.02,
  "proj_mlp": true,
  "proj_hidden_mult": 4,

  "lambda_lm": 1.0,
  "lambda_id": 1.0,
  "lambda_kl": 0.0001,
  "tau_infonce": 0.2,
  "infonce_full_answer": true,
  "infonce_target_max_len": 128,

  "lr_lora": 2e-4,
  "lr_proj": 1e-4,
  "lr_head": 3e-4,
  "weight_decay": 0.01,
  "max_grad_norm": 1.0,
  "warmup_steps": 200,

  "batch_size": 4,
  "grad_accum": 4,
  "max_steps": 3500,
  "max_prompt_len": 192,
  "max_answer_len": 192,

  "log_every": 10,
  "eval_every": 0,
  "eval_size": 200,
  "save_every": 1000,
  "seed": 42,

  "output_dir": "/home/ubuntu/work/blt_exp7b_opt13",
  "data_train_size": null,
  "data_eval_size": 200
}