Practical notes: Building LLM, Part 1 — Training: Predicting the Next Token

Use this as an operator-facing rebuild of the ideas in “Building LLM, Part 1 — Training: Predicting the Next Token”: clear stages, ordered code slots, and recovery notes that survive a handoff. The Overview stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Why a Vision Transformer first?

For the Why a Vision Transformer stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

The one idea: predict the next token

For the The one idea predict stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

logits, _ = model(ids)                     # ids = the tokens so far, shape (1, T)
probs = F.softmax(logits[0, -1], dim=-1)   # one probability per vocabulary token

Text → tokens: why we had to build our own tokenizer

For the Text tokens why we stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call. For the Text tokens why we stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

import glob
from tokenizers import Tokenizer, models, trainers, pre_tokenizers
from llm_transformer.prep_tales import strip_gutenberg, RAW_DIR

# read every raw .txt book and strip its Project Gutenberg header/footer,
# leaving just the story text - one big string per book
tales_texts = [strip_gutenberg(open(p, encoding="utf-8", errors="ignore").read())
               for p in sorted(glob.glob(f"{RAW_DIR}/*.txt"))]   # 12 books, ~809k words
tok = Tokenizer(models.BPE())
tok.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)   # start from bytes
trainer = trainers.BpeTrainer(vocab_size=8192, special_tokens=["<|endoftext|>"],
                              initial_alphabet=pre_tokenizers.ByteLevel.alphabet())
tok.train_from_iterator(tales_texts, trainer)   # learn the 7,935 merges
tok.save("tales_bpe.json")

tok = Tokenizer.from_file("tales_bpe.json")   # reload the trained tokenizer
ids = tok.encode("Once upon a time").ids      # -> [412, 987, 15, 733]   (integers)
idx = torch.tensor(ids).unsqueeze(0)          # (1, 4) — a batch of one sequence

vocab_size = tok.get_vocab_size()             # 8192 - sets the NUMBER OF ROWS below

Token → vector: a lookup, not a computation

When working through the Token vector a lookup stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn.

self.token_embed = nn.Embedding(vocab_size, n_embed)   # the 8,192 x 256 table
# .weight IS a learnable (8192, 256) parameter tensor — 2.1M weights, trained by backprop;
# each id's row is nudged only on steps where that token appears in the batch
...
tok_emb = self.token_embed(idx)     # idx (B, T)  ->  vectors (B, T, 256), one row per id

The exception: the causal mask

When working through the The exception the causal stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn.

# tril = torch.tril(torch.ones(T, T))  — a lower-triangular matrix of 1s, made once
attn = q @ k.transpose(-2, -1) / math.sqrt(head_size)          # scores, (B, T, T)
attn = attn.masked_fill(self.tril[:T, :T] == 0, float('-inf')) # blank out the future
attn = F.softmax(attn, dim=-1)                                  # -inf -> weight 0
out  = attn @ v

One pass, every context at once

When working through the One pass every context stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn. When working through the One pass every context stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Inside one run: what attn @ V actually computes

The Inside one run what stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

One pass, many lessons: T predictions, T losses

The One pass many lessons stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices.

logits = self.lm_head(x)                       # (B, T, vocab)
B, T, C = logits.shape
loss = F.cross_entropy(logits.reshape(B*T, C), # every position ...
                       targets.reshape(B*T))   # ... vs the token that actually followed

Each loss is a vocabulary classification (just like the ViT, but 8,192 classes)

The Each loss is a stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Budget tokens per turn and per session. Agentic tools expand context aggressively; hard caps keep demos from becoming surprise invoices. The Each loss is a stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

How a batch is built — and how its one loss is formed

For the How a batch is stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

Training: watching the loss come down

For the Training watching the loss stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

for step in range(train_steps):
    xb, yb = get_batch(train_data)     # random (context, next-token) windows
    _, loss = model(xb, yb)            # forward: T predictions -> one mean loss
    opt.zero_grad(); loss.backward()   # backprop the loss into all 9M parameters
    opt.step()                         # nudge them downhill

What we built — and what’s next

For the What we built and stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call. For the What we built and stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Operational checklist

When working through the Operational checklist stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest.

Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn.

Keep render work cheap and push expensive derivation behind memoization only after measuring. Premature memo can hide stale props bugs.

Add a smoke test that exercises the critical path in CI with fixtures, not live paid APIs, whenever budgets allow.

Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Before promoting the stack, freeze versions, capture a golden transcript for the critical path, and confirm rollback steps. Shared environments need rate limits, tenancy checks, and a clear owner for secret rotation. Prefer boring reliability over clever one-off demos.

Batch note for 44c72473e486: keep provider keys out of the repo, set a per-session token ceiling, and store transcripts next to the eval fixtures so later model swaps stay comparable.

For the hardening note 0 stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

Hardening detail 0/729: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

When working through the hardening note 1 stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.

Hardening detail 1/729: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

The hardening note 2 stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

Hardening detail 2/729: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

For the hardening note 3 stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Hardening detail 3/729: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

When working through the hardening note 4 stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Hardening detail 4/729: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

The hardening note 5 stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

Hardening detail 5/729: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

For the hardening note 6 stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.

Hardening detail 6/729: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

When working through the hardening note 7 stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline.

Hardening detail 7/729: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

The hardening note 8 stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Hardening detail 8/729: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

For the hardening note 9 stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Hardening detail 9/729: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

When working through the hardening note 10 stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

Hardening detail 10/729: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.

The hardening note 11 stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.

Hardening detail 11/729: measure wall time, error class, and token spend for this note, then decide whether to keep the change based on a fixed question set rather than anecdote.