sft_rag_data.py

"""Build retrieval-augmented SFT data.

Two skills mixed:
  (A) GROUNDED answering from context (SberQuAD): teaches "use the provided context".
      prompt = "Контекст: {ctx}\nВопрос: {q}\nОтвет:"  ; answer = " {span}"<eos>
  (B) GENERAL instruction following (saiga_scored, no context), to keep fluency.
      prompt = "Вопрос: {q}\nОтвет:"                    ; answer = " {bot}"<eos>
Loss only on answer (prompt+pad = -100). Saves /root/sft_rag.pt.
"""
import torch, random
from datasets import load_dataset
from tokenizers import Tokenizer

TOK = "/root/ru_tok.json"
OUT = "/root/sft_rag.pt"
MAXLEN = 512
MINSCORE = 8

tok = Tokenizer.from_file(TOK)
BOS = tok.token_to_id("<bos>")
EOS = tok.token_to_id("<eos>")


def enc(s):
    return tok.encode(s).ids


examples = []

# (A) SberQuAD — grounded
sq = load_dataset("sberquad", split="train")
pre = enc("Контекст: ")
n_ctx_trunc = 0
for r in sq:
    ctx, q = r["context"].strip(), r["question"].strip()
    ans = r["answers"]["text"][0].strip() if r["answers"]["text"] else ""
    if not ctx or not q or not ans:
        continue
    midq = enc("\nВопрос: " + q + "\nОтвет:")
    a_ids = enc(" " + ans) + [EOS]
    budget = MAXLEN - 1 - len(pre) - len(midq) - len(a_ids)
    if budget < 32:
        continue
    c_ids = enc(ctx)
    if len(c_ids) > budget:
        c_ids = c_ids[:budget]; n_ctx_trunc += 1
    ids = [BOS] + pre + c_ids + midq + a_ids
    labels = [-100] * (len(ids) - len(a_ids)) + a_ids
    examples.append({"ids": ids, "labels": labels})
n_sq = len(examples)

# (B) saiga_scored — general, no context
sa = load_dataset("IlyaGusev/saiga_scored", split="train")
for r in sa:
    if not (r["language"] == "Russian" and not r["is_bad_by_regex"]
            and (r["opus_score"] or 0) >= MINSCORE and r["turns"] <= 1
            and len(r["messages"]) >= 2
            and r["messages"][0]["role"] == "user"
            and r["messages"][1]["role"] == "bot"):
        continue
    u = r["messages"][0]["content"].strip()
    b = r["messages"][1]["content"].strip()
    if not u or not b:
        continue
    p_ids = enc(f"Вопрос: {u}\nОтвет:")
    a_ids = enc(" " + b) + [EOS]
    ids = [BOS] + p_ids + a_ids
    if len(ids) > MAXLEN:
        continue
    labels = [-100] * (1 + len(p_ids)) + a_ids
    examples.append({"ids": ids, "labels": labels})
n_sa = len(examples) - n_sq

random.seed(0)
random.shuffle(examples)
nval = 1000
val, train = examples[:nval], examples[nval:]
torch.save({"train": train, "val": val}, OUT)
ntok = sum(len(e["ids"]) for e in examples)
print(f"sberquad={n_sq} (ctx_trunc={n_ctx_trunc}) saiga={n_sa} total={len(examples)} "
      f"train={len(train)} val={len(val)} tokens={ntok/1e6:.2f}M avg_len={ntok/len(examples):.0f}")
print("saved", OUT)