File size: 12,469 Bytes
4968ea3 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 | """Pseudo-token span utilities (NO vocab extension).
Control signals ([MS:*] [ACT:*] [EOQ] [ANS] [RETRIEVED]) are plain text strings
fed through the original Qwen3-8B tokenizer's subword splitting. We never call
add_special_tokens / resize embeddings / touch lm_head. This is the Search-R1 /
R1 route: the framework parses control strings deterministically and masks token
spans by *segmented* tokenization + length-based concatenation (NOT substring β
token-boundary guessing, which is ambiguous).
This module provides three pure helpers:
- tokenize_with_spans: split a target into [MS:*] / [ACT:*]..[EOQ] / rest segments,
encode each with add_special_tokens=False, concatenate, and return target_ids
plus two token-level bool masks (ms_mask, act_query_mask).
- find_marker_token_span: locate a marker's token span inside an id list
(used on the rollout side to cut the [EOQ] boundary and locate the [MS:*] span).
- extract_control_strings: regex out [MS:X], [ACT:Y], <rewrite_query> from text;
malformed output returns a fallback flag. reward/parse use strings, not ids.
"""
import re
from typing import List, Optional, Tuple
from src.utils.special_tokens import (
ACT_TOKENS,
EOQ_TOKEN,
MS_TOKENS,
)
# Regex anchors over the pseudo-token strings. We allow optional surrounding
# whitespace; the strings themselves are literal "[MS:SM]" etc.
_MS_RE = re.compile(r"\[MS:(SM|PM|VM|NM)\]")
_ACT_RE = re.compile(r"\[ACT:(DIRECT|REWRITE|CLUE|RETRIEVE)\]")
_EOQ_RE = re.compile(re.escape(EOQ_TOKEN))
_VALID_MS = set(MS_TOKENS.keys())
_VALID_ACT = set(ACT_TOKENS.keys())
def tokenize_with_spans(
tok,
prompt_str: str,
target_str: str,
) -> dict:
"""Tokenize prompt + target and return target spans for loss weighting.
The target string is expected to look like:
"[MS:PM] [ACT:REWRITE] <rewrite query text> [EOQ]"
or for SM/DIRECT:
"[MS:SM] [ACT:DIRECT] [EOQ]" (empty rewrite span)
We segment the target into three pieces so we can label spans precisely:
seg_ms : the "[MS:*]" marker (+ trailing space)
seg_act_q: "[ACT:*] <query> [EOQ]" (the act marker through and including EOQ)
(there is no "rest" after EOQ in SFT targets; everything is covered)
Each segment is encoded with add_special_tokens=False and concatenated. The
masks are built from segment token lengths (no substring search), which is the
robust Search-R1 alignment trick.
Returns dict with:
prompt_ids: List[int]
target_ids: List[int]
ms_mask: List[bool] len == len(target_ids); True over [MS:*] subwords
act_query_mask: List[bool] len == len(target_ids); True over [ACT:*]..[EOQ]
ms_str: the matched "[MS:*]" string (or None)
act_str: the matched "[ACT:*]" string (or None)
"""
prompt_ids = tok.encode(prompt_str, add_special_tokens=False)
ms_match = _MS_RE.search(target_str)
act_match = _ACT_RE.search(target_str)
eoq_idx = target_str.find(EOQ_TOKEN)
target_ids: List[int] = []
ms_mask: List[bool] = []
act_query_mask: List[bool] = []
def _append(segment: str, is_ms: bool, is_act_q: bool):
if segment == "":
return
seg_ids = tok.encode(segment, add_special_tokens=False)
target_ids.extend(seg_ids)
ms_mask.extend([is_ms] * len(seg_ids))
act_query_mask.extend([is_act_q] * len(seg_ids))
if ms_match is None or act_match is None or eoq_idx == -1:
# Malformed target (should not happen for SFT-built targets). Encode whole
# thing as a generic (unmasked) span so training still proceeds.
_append(target_str, is_ms=False, is_act_q=False)
return {
"prompt_ids": prompt_ids,
"target_ids": target_ids,
"ms_mask": ms_mask,
"act_query_mask": act_query_mask,
"ms_str": ms_match.group(0) if ms_match else None,
"act_str": act_match.group(0) if act_match else None,
}
# Segment boundaries (character offsets in target_str):
# [0, ms_start) : any leading text (usually empty) -> generic
# [ms_start, ms_end) : [MS:*] -> ms span
# [ms_end, act_start) : between MS and ACT (whitespace) -> generic
# [act_start, eoq_end) : [ACT:*] .. [EOQ] (inclusive) -> act_query span
# [eoq_end, end) : trailing (should be empty) -> generic
ms_start, ms_end = ms_match.start(), ms_match.end()
act_start = act_match.start()
eoq_end = eoq_idx + len(EOQ_TOKEN)
_append(target_str[:ms_start], is_ms=False, is_act_q=False)
_append(target_str[ms_start:ms_end], is_ms=True, is_act_q=False)
_append(target_str[ms_end:act_start], is_ms=False, is_act_q=False)
_append(target_str[act_start:eoq_end], is_ms=False, is_act_q=True)
_append(target_str[eoq_end:], is_ms=False, is_act_q=False)
assert len(ms_mask) == len(act_query_mask) == len(target_ids)
return {
"prompt_ids": prompt_ids,
"target_ids": target_ids,
"ms_mask": ms_mask,
"act_query_mask": act_query_mask,
"ms_str": ms_match.group(0),
"act_str": act_match.group(0),
}
def tokenize_with_answer(
tok,
prompt_str: str,
decision_segments: List[Tuple[str, str]],
rag_block: str,
answer_suffix: str,
answer_str: str,
eos_id: Optional[int],
model_max: Optional[int] = None,
) -> dict:
"""Phase-C SFT tokenization: prompt + decision + retrieved block + answer cue + answer.
π΄ This MUST be byte-identical (token-id level) to the rollout's teacher-forcing prefix
(src/train/rl/rollout.py). The rollout encodes each segment SEPARATELY with
add_special_tokens=False (Qwen BPE merges across segment boundaries if concatenated
first), so we replicate the SAME per-segment encode order. The decision segment encode
differs PER forced branch, so 08c pre-splits it into `decision_segments` (a list of
(seg_str, role) where role β {"dec", "query_fixed", "query_content"}); we just encode
each seg_str in order β NOT reuse tokenize_with_spans (whose [MS]/space/[ACT]..[EOQ]
split produces different ids). See plan critique #2.
Segment order (mirrors rollout):
prompt_str β label -100, answer_mask False
*decision_segments β label real, answer_mask False (teacher-forced context)
rag_block ("" for SM β skip)β label -100, answer_mask False
answer_suffix (ends [ANS]) β label -100, answer_mask False
answer_str β label real, answer_mask True
eos β label real, answer_mask True π΄ eos ONLY here
π΄ eos is appended ONLY after the final answer, NEVER after the decision segment
(the existing tokenize_with_spans path appends eos after the decision target; doing
that here would make the Phase-C prefix diverge from the rollout β critique #4).
Phase C trains ONLY the answer span (weighted_ce_loss active = labels!=-100 &
answer_mask), so decision labels being "real" is harmless β they never enter the loss.
Returns dict: input_ids / labels / answer_mask (equal-length List[int]/List[int]/List[bool]).
"""
input_ids: List[int] = []
labels: List[int] = []
answer_mask: List[bool] = []
def _append(segment: str, is_label: bool, is_answer: bool):
if segment == "":
return
seg_ids = tok.encode(segment, add_special_tokens=False)
input_ids.extend(seg_ids)
labels.extend(seg_ids if is_label else [-100] * len(seg_ids))
answer_mask.extend([is_answer] * len(seg_ids))
_append(prompt_str, is_label=False, is_answer=False)
for seg_str, _role in decision_segments:
_append(seg_str, is_label=True, is_answer=False)
_append(rag_block, is_label=False, is_answer=False) # "" for SM/DIRECT β skipped
_append(answer_suffix, is_label=False, is_answer=False)
_append(answer_str, is_label=True, is_answer=True)
if eos_id is not None:
input_ids.append(eos_id)
labels.append(eos_id)
answer_mask.append(True)
if model_max is not None and len(input_ids) > model_max:
cut = len(input_ids) - model_max
input_ids = input_ids[cut:]
labels = labels[cut:]
answer_mask = answer_mask[cut:]
assert len(input_ids) == len(labels) == len(answer_mask)
return {"input_ids": input_ids, "labels": labels, "answer_mask": answer_mask}
def find_marker_token_span(
ids: List[int],
tok,
marker_str: str,
start: int = 0,
) -> Optional[Tuple[int, int]]:
"""Find the first token span in ids whose decoded text contains marker_str.
Decode-based (NOT exact id-subsequence): a marker's subword tokenization differs
by what precedes it (e.g. "[EOQ]" -> [58,6760,48,60] standalone vs [508,6760,48,60]
when glued to a preceding space). So we incrementally decode and locate the marker
string. Returns (begin, end) β token indices such that ids[:end] is the smallest
prefix whose tail contains marker_str, and ids[begin:end] is the tightest window
still containing it. Returns None if not found.
Used on the rollout side to cut the decision segment right after the first [EOQ]
(keep ids[:end]) and to locate the [MS:*] span. O(n^2) decodes but n is tiny
(<=64 here). [ANS] is NOT searched β it is concatenated as a known-length suffix.
"""
n = len(ids)
for end in range(start + 1, n + 1):
if marker_str in tok.decode(ids[start:end]):
# tighten the begin: largest begin < end still containing the marker
begin = start
for b in range(start, end):
if marker_str in tok.decode(ids[b:end]):
begin = b
else:
break
return (begin, end)
return None
# control pseudo-tokens that must NEVER appear in a final answer string. They are plain
# text (not registered special tokens), so tokenizer.decode(skip_special_tokens=True)
# does NOT remove them β the model, trained to emit [EOQ] in the decision stage, often
# tails the answer with "... [EOQ]". We truncate at the first such marker and strip any
# leftovers, so eval EM/F1 and DAPO reward (both read traj.answer) are not polluted.
_ANSWER_STOP_MARKERS = ("[EOQ]", "[ANS]", "[RETRIEVED]", "[/RETRIEVED]")
_CONTROL_TOKEN_RE = re.compile(r"\[(?:MS|ACT):[A-Z]+\]|\[/?(?:EOQ|ANS|RETRIEVED|QUERY)\]")
def clean_answer_text(text: str) -> str:
"""Strip control pseudo-tokens from a generated answer.
Truncate at the FIRST answer-stop marker ([EOQ]/[ANS]/[RETRIEVED]/[/RETRIEVED]) β
anything the model emits after it is control garbage, not answer content β then
remove any stray [MS:*]/[ACT:*]/[EOQ]/... tokens that slipped in earlier, and strip
whitespace. Idempotent; a clean answer passes through unchanged.
"""
cut = len(text)
for m in _ANSWER_STOP_MARKERS:
i = text.find(m)
if i != -1:
cut = min(cut, i)
text = text[:cut]
text = _CONTROL_TOKEN_RE.sub("", text)
return text.strip()
def extract_control_strings(text: str) -> dict:
"""Parse [MS:X], [ACT:Y], and the rewrite_query span from generated text.
The rewrite_query is the text between [ACT:*] and [EOQ] (stripped). For
DIRECT the span is typically empty.
Returns dict:
ms: one of SM/PM/VM/NM or None
act: one of DIRECT/REWRITE/CLUE/RETRIEVE or None
query_text: str (may be "")
valid: bool β True iff both [MS:*] and [ACT:*] were found
has_eoq: bool β whether an [EOQ] terminator was present
"""
ms_match = _MS_RE.search(text)
act_match = _ACT_RE.search(text)
ms = ms_match.group(1) if ms_match else None
act = act_match.group(1) if act_match else None
query_text = ""
has_eoq = False
if act_match is not None:
after_act = text[act_match.end():]
eoq_pos = after_act.find(EOQ_TOKEN)
if eoq_pos != -1:
has_eoq = True
query_text = after_act[:eoq_pos].strip()
else:
query_text = after_act.strip()
return {
"ms": ms,
"act": act,
"query_text": query_text,
"valid": (ms is not None and act is not None),
"has_eoq": has_eoq,
}
|