原书出处
Build a Reasoning Model (From Scratch)
f"train_rlvr_grpo_metrics_{timestamp}.csv" csv_log_path = Path(csv_log_path) try: #B for step in range(steps): #C optimizer.zero_grad() current_step = step + 1 example = math_data[step % len(math_data)] #D stats = compute_grpo_loss( model=model, tokenizer=tokenizer, example=example, device=device, num_rollouts=num_rollouts, max_new_tokens=max_new_tokens, temperature=temperature, top_p=top_p, ) #E stats["loss_tensor"].backward() #F torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0) #G optimizer.step() #H reward_avg = torch.tensor(stats["rewards"]).mean().item() step_tokens = sum( sample["gen_len"] for sample in stats["samples"] ) avg_response_len = ( step_tokens / len(stats["samples"]) if stats["samples"] else 0.0
原书定位详情
来源:a19c7933-4eda-4176-aebd-30fe7e1ad292
正文位置:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-6.html
EPUB SHA-256:52d4d07986020be70f5aac21762dfcf60fcba08db0ba29d4f63957caf12aa377