LLM Training & Generation · 训练与生成入门

LLM Training and Generation: Small Implementation Exercises

PaperLeet original curriculum; primary documentation linked per exercise

Read source material

From next-token targets and response masks to candidate probabilities, token-weighted accumulation and global clipping.

What you will implement

Seven standalone exercises with explicit shifting, masks, ties and reductions; deterministic probability experiments and actual small-model gradient updates.

• 86. Next-token Loss: Shifted Causal Targets

  Align predictions with the next token and compute a stable average loss.

  Implement one core function; independent tests and deterministic synthetic experiments are provided.

• 87. SFT Loss Mask: Train on Response Tokens

  Shift targets, mask prompt and padding tokens, and average loss over valid response tokens.

  Implement one core function; independent tests and deterministic synthetic experiments are provided.

• 94. Multi-turn SFT: Assistant Targets and Conversation Boundaries

  Select assistant targets across turns while excluding padding and cross-conversation predictions.

  Implement one boolean mask; roles, conversation IDs and truncated sequences are supplied. No tokenizer, chat template or attention-mask construction.

• 88. Top-k: Temperature and Candidate Probabilities

  Apply temperature, keep exactly k candidates, and renormalize their probabilities.

  Implement one core function; independent tests and deterministic synthetic experiments are provided.

• 89. Top-p: Cumulative Probability and Nucleus Candidates

  Keep the smallest candidate set that reaches the probability threshold, then restore vocabulary order.

  Implement one core function; independent tests and deterministic synthetic experiments are provided.

• 90. Gradient Accumulation: Valid-token Weighting

  Weight microbatch gradients by valid token counts to match one combined batch.

  Implement one core function; independent tests and deterministic synthetic experiments are provided.

• 91. Global Gradient Clipping: One Norm for All Parameters

  Compute one gradient norm across all parameters and apply a shared clipping scale.

  Implement one core function; independent tests and deterministic synthetic experiments are provided.