Capstone A: fine-tune SmolLM2 end-to-end, by hand

This capstone assembles Lessons 2.1–2.8 into one script you can run top to bottom. Nothing here is new — it's the whole loop in one place, so you can see how few moving parts a real fine-tune has once you understand each one.

The whole pipeline, one script

import torch
from datasets import Dataset
from transformers import (AutoModelForCausalLM, AutoTokenizer,
                          TrainingArguments, Trainer, DataCollatorForSeq2Seq)
from peft import LoraConfig, get_peft_model, PeftModel

MODEL_ID, MAX_LEN = "HuggingFaceTB/SmolLM2-135M-Instruct", 256
device = "cuda" if torch.cuda.is_available() else "cpu"

# 1) data --------------------------------------------------------------
raw = [  # in practice: hundreds of clean, balanced, deduplicated rows
    {"prompt": "Classify the sentiment as positive or negative: I loved this movie.", "completion": "positive"},
    {"prompt": "Classify the sentiment as positive or negative: A complete waste of time.", "completion": "negative"},
    # ...
]
ds = Dataset.from_list(raw).train_test_split(test_size=0.2, seed=42)

# 2) model + tokenizer -------------------------------------------------
tok = AutoTokenizer.from_pretrained(MODEL_ID)
if tok.pad_token is None: tok.pad_token = tok.eos_token
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, dtype=torch.bfloat16).to(device)
model.config.pad_token_id = tok.pad_token_id

# 3) tokenize with a loss mask ----------------------------------------
def tokenize(row):
    full = tok.apply_chat_template(
        [{"role":"user","content":row["prompt"]},{"role":"assistant","content":row["completion"]}], tokenize=False)
    prompt = tok.apply_chat_template(
        [{"role":"user","content":row["prompt"]}], tokenize=False, add_generation_prompt=True)
    ids = tok(full, truncation=True, max_length=MAX_LEN, add_special_tokens=False)["input_ids"]
    p = len(tok(prompt, add_special_tokens=False)["input_ids"])
    labels = [-100]*min(p, len(ids)) + ids[min(p, len(ids)):]
    return {"input_ids": ids, "attention_mask":[1]*len(ids), "labels": labels}
train_tok = ds["train"].map(tokenize, remove_columns=ds["train"].column_names)
val_tok   = ds["test"].map(tokenize, remove_columns=ds["test"].column_names)

# 4) LoRA + train ------------------------------------------------------
model = get_peft_model(model, LoraConfig(
    r=16, lora_alpha=32, lora_dropout=0.05,
    target_modules=["q_proj","v_proj"], task_type="CAUSAL_LM"))
model.print_trainable_parameters()
Trainer(
    model=model,
    args=TrainingArguments(output_dir="sft-out", per_device_train_batch_size=8,
        gradient_accumulation_steps=2, learning_rate=2e-4, num_train_epochs=3,
        lr_scheduler_type="cosine", warmup_ratio=0.03, bf16=True,
        logging_steps=5, eval_strategy="epoch", save_strategy="epoch", report_to=[]),
    train_dataset=train_tok, eval_dataset=val_tok,
    data_collator=DataCollatorForSeq2Seq(tok, model=model), processing_class=tok,
).train()
model.save_pretrained("sft-out/adapter")

# 5) evaluate vs base on a gold set -----------------------------------
@torch.no_grad()
def predict(m, prompt):
    text = tok.apply_chat_template([{"role":"user","content":prompt}], tokenize=False, add_generation_prompt=True)
    ids = tok(text, return_tensors="pt").to(m.device)
    return tok.decode(m.generate(**ids, max_new_tokens=8, do_sample=False, pad_token_id=tok.pad_token_id)[0][ids["input_ids"].shape[1]:], skip_special_tokens=True).strip().lower()

gold = [{"prompt":"Classify the sentiment as positive or negative: This made my week.","label":"positive"}]  # + more
base = AutoModelForCausalLM.from_pretrained(MODEL_ID, dtype=torch.bfloat16).to(device)
tuned = PeftModel.from_pretrained(base, "sft-out/adapter").eval()
acc = lambda m: sum(predict(m, g["prompt"]).startswith(g["label"]) for g in gold) / len(gold)
print("base:", acc(base), " tuned:", acc(tuned))

# 6) ship --------------------------------------------------------------
tuned.merge_and_unload().save_pretrained("sft-out/merged"); tok.save_pretrained("sft-out/merged")

Success criteria

You've succeeded when, on your held-out gold set, the tuned model clearly beats the base and produces the format you want (here: just the label, no rambling). If you set a gate in advance (Track 1) — say "≥ 90% gold accuracy" — the run passes only when it clears it.

You can fine-tune a model

That's the whole thing. You loaded a base model, taught it a task from your own data, proved the lift on a gold set, and shipped an artifact — entirely by hand, with no platform and nothing hidden. Everything BrewSLM does, you can now do yourself; you understand each step well enough to debug it.

So why use a platform at all? Because doing this once by hand is illuminating; doing it fifty times — with imports, gates, eval packs, deployment, and monitoring — is exactly the toil worth automating. Track 3 runs this same pipeline through BrewSLM and shows what the platform buys you, now that you know precisely what it's doing underneath.