File size: 7,920 Bytes
a962f7c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 | # sft.py
import json, math, random, argparse
from pathlib import Path
import torch, torch.nn as nn
from datasets import load_dataset
from main import Transformer
from bpe import load_bpe_tokenizer
import os
import re
SCRIPT_DIR = Path(__file__).parent.absolute()
ROOT = SCRIPT_DIR.parent # .../Alpha
DATA = ROOT / "jsons" # search_ex, calc_ex, rag_*.json
TOK = ROOT / "jsons" / "fineweb_bpe.json" # confirm with the find command
CKPT = ROOT / "checkpoints"
MAX_LEN, MICRO, ACCUM = 2048, 16, 2
LR, WARMUP, EPOCHS, EVAL_EVERY = 3e-5, 50, 3, 200
def fmt(ex):
if ex.get("input"):
return (f"### Instruction:\n{ex['instruction']}\n\n"
f"### Input:\n{ex['input']}\n\n### Response:\n")
return f"### Instruction:\n{ex['instruction']}\n\n### Response:\n"
def build(ds, tok):
bos, eos = tok.word_to_idx["<bos>"], tok.word_to_idx["<eos>"]
enc = lambda s: tok.tok.encode(s, add_special_tokens=False).ids
out = []
for ex in ds:
if "messages" in ex:
item = encode_chat(ex, bos, eos, enc)
if item: out.append(item)
continue
p = [bos] + enc(fmt(ex))
a = enc(ex["output"].strip()) + [eos]
if len(p) + len(a) > MAX_LEN or len(a) < 2:
continue # drop, never truncate: a cut answer would lose its <eos>
out.append((p + a, [-100] * len(p) + a))
return out
def encode_chat(ex, bos, eos, enc):
msgs, sysm = ex["messages"], ""
if msgs[0]["role"] == "system":
sysm, msgs = msgs[0]["content"], msgs[1:]
ids, lab = [bos], [-100]
for i in range(0, len(msgs) - 1, 2):
head = f"### System:\n{sysm}\n\n" if (i == 0 and sysm) else ""
p = enc(f"{head}### Instruction:\n{msgs[i]['content']}\n\n### Response:\n")
r = enc(msgs[i + 1]["content"]) + [eos]
if len(ids) + len(p) + len(r) > MAX_LEN:
break # stop at a turn boundary, never mid-answer
ids += p + r
lab += [-100] * len(p) + r # loss only on assistant text + <eos>
return (ids, lab) if len(ids) > 1 else None
def collate(batch, pad):
n = max(len(x) for x, _ in batch)
ids = torch.full((len(batch), n), pad)
lab = torch.full((len(batch), n), -100)
for i, (x, y) in enumerate(batch):
ids[i, :len(x)] = torch.tensor(x)
lab[i, :len(y)] = torch.tensor(y)
return ids[:, :-1], lab[:, 1:] # right-padding is safe: causal attention
def lr_at(s, total):
if s < WARMUP: return LR * (s + 1) / WARMUP
p = (s - WARMUP) / max(1, total - WARMUP)
return 0.1 * LR + 0.5 * 0.9 * LR * (1 + math.cos(math.pi * p))
@torch.no_grad()
def evaluate(model, data, pad, dev):
model.eval(); tot, cnt = 0.0, 0
for i in range(0, len(data), MICRO):
x, y = collate(data[i:i + MICRO], pad)
x, y = x.to(dev), y.to(dev)
with torch.autocast("cuda", dtype=torch.bfloat16, enabled=dev.type == "cuda"):
lg = model(x)
tot += nn.functional.cross_entropy(lg.float().reshape(-1, lg.size(-1)), y.reshape(-1),
ignore_index=-100, reduction="sum").item()
cnt += (y != -100).sum().item()
model.train()
return tot / cnt
if __name__ == "__main__":
ap = argparse.ArgumentParser()
ap.add_argument("--base", default=str(CKPT / "checkpoint_latest.pt"))
ap.add_argument("--out", default=str(CKPT / "checkpoint_sftv4.pt"))
a = ap.parse_args()
Path(a.out).parent.mkdir(parents=True, exist_ok=True)
random.seed(42); torch.manual_seed(42)
dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
tok = load_bpe_tokenizer(TOK)
pad = tok.word_to_idx["<pad>"]
alpaca = list(load_dataset("yahma/alpaca-cleaned")["train"])
MATH = re.compile(r"^(calculate|compute|evaluate|solve|add|subtract|multiply|divide|"
r"find the (sum|product|difference|quotient|result)|"
r"what is the (sum|product|difference|quotient|result)|"
r"what('s| is) [\d\.\-\+\*/%\s]+)", re.I)
alpaca = [ex for ex in alpaca if ex["input"] or not MATH.match(ex["instruction"].strip())]
FACT = re.compile(r"^(who|when|where|which|what is|what was|what are|how many|how much|in which|name the)\b", re.I)
alpaca = [ex for ex in alpaca
if ex["input"] or not FACT.match(ex["instruction"].strip())]
with open(DATA / "search_ex.json", encoding="utf-8") as f:
search = json.load(f)
with open(DATA / "calc_ex.json", encoding="utf-8") as f:
calc = json.load(f)
with open(DATA / "rag_ex.json", encoding="utf-8") as f:
rag = json.load(f)
with open(DATA / "rag_long.json", encoding="utf-8") as f:
rag_long = json.load(f)
with open(DATA / "rag_asqa.json", encoding="utf-8") as f:
rag_asqa = json.load(f)
with open(DATA / "rag_wiki.json", encoding="utf-8") as f:
rag_wiki = json.load(f)
with open(DATA / "chat.json", encoding="utf-8") as f:
chat = json.load(f)
with open(DATA / "identity.json", encoding="utf-8") as f:
identity = json.load(f)
def first_user(ex):
return next(m["content"] for m in ex["messages"] if m["role"] == "user").strip()
chat = [ex for ex in chat if not (MATH.match(first_user(ex)) or FACT.match(first_user(ex)))]
raw = alpaca + search + calc + rag + rag_long * 2 + rag_asqa + rag_wiki + chat + identity * 3
random.shuffle(raw)
data = build(raw, tok)
val, train = data[:1000], data[1000:]
print(f"train {len(train)} | val {len(val)} examples")
ck = torch.load(a.base, map_location="cpu"); arch = ck["arch"]
model = Transformer(arch["d_model"], arch["num_heads"], arch["vocab_size"],
arch["d_ff"], arch["num_kv_heads"], arch["num_decoder_layers"])
model.load_state_dict(ck["model_state_dict"]); model.to(dev).train()
opt = torch.optim.AdamW(model.parameters(), lr=LR, betas=(0.9, 0.95),
weight_decay=0.0, fused=dev.type == "cuda")
per_step = MICRO * ACCUM
total = EPOCHS * (len(train) // per_step)
print(f"base val (answer tokens): {evaluate(model, val, pad, dev):.4f}")
step, best = 0, float("inf")
for ep in range(EPOCHS):
random.shuffle(train)
for b in range(0, len(train) - per_step + 1, per_step):
for g in opt.param_groups: g["lr"] = lr_at(step, total)
opt.zero_grad(set_to_none=True); tl = 0.0
for m in range(ACCUM):
x, y = collate(train[b + m * MICRO: b + (m + 1) * MICRO], pad)
x, y = x.to(dev), y.to(dev)
with torch.autocast("cuda", dtype=torch.bfloat16, enabled=dev.type == "cuda"):
lg = model(x)
loss = nn.functional.cross_entropy(lg.float().reshape(-1, lg.size(-1)),
y.reshape(-1), ignore_index=-100)
(loss / ACCUM).backward(); tl += loss.item() / ACCUM
gn = nn.utils.clip_grad_norm_(model.parameters(), 1.0)
opt.step(); step += 1
if step % 20 == 0:
print(f"ep {ep+1} step {step}/{total} | loss {tl:.4f} | gnorm {gn:.2f}")
if step % EVAL_EVERY == 0 or step == total:
v = evaluate(model, val, pad, dev)
print(f"[Eval] step {step} | val {v:.4f}")
if v < best:
best = v
torch.save({"model_state_dict": model.state_dict(), "arch": arch,
"step": step, "val_loss": v}, a.out)
print(" saved best") |