File size: 2,424 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 | """Shared context-budget / truncation helpers (baseline-identical numbers).
🔴 RL rollout AND eval must use the SAME budget (plan §4.3 / §4.4 / eval):
- model_max = 32000, gen_length(answer) = 800
- max_retrieval_tokens = 32000 - 800 - 1000 = 30200
- retrieved-session text exceeding 30200 → char-ratio crop (baseline build_prompt_rag)
- final full_ids exceeding 32000 → LEFT-truncate (truncation_side="left"): drop the
earliest retrieved text first, keep recent decision/answer markers.
- answer max_new_tokens = 800 (RL and eval IDENTICAL — never 200 here / 800 there)
- decision segment max_new_tokens = 64 (only [MS:*][ACT:*]<query>[EOQ])
estimate_tokens uses len//4 (baseline). The char-ratio crop mirrors baseline exactly.
"""
from typing import List, Tuple
MODEL_MAX = 32000
GEN_LENGTH = 800
RESERVE = 1000
MAX_RETRIEVAL_TOKENS = MODEL_MAX - GEN_LENGTH - RESERVE # 30200
DECISION_MAX_NEW_TOKENS = 64
ANSWER_MAX_NEW_TOKENS = GEN_LENGTH # 800
def estimate_tokens(text: str) -> int:
return len(text) // 4
def crop_retrieval_text(history_string: str, max_tokens: int = MAX_RETRIEVAL_TOKENS) -> str:
"""Char-ratio crop of the assembled retrieved-session string (baseline口径)."""
est = estimate_tokens(history_string)
if est > max_tokens:
ratio = max_tokens / est
char_limit = int(len(history_string) * ratio)
return history_string[:char_limit]
return history_string
def left_truncate_ids(full_ids: List[int], mask: List[bool],
max_length: int = MODEL_MAX,
*extra_masks: List[bool]):
"""LEFT-truncate full_ids (+ aligned mask and any extra aligned masks) to max_length.
Drops the earliest tokens (retrieved text sits at the front of the post-decision
region; the decision prompt itself also sits at the very front). This matches
baseline tokenizer(truncation=True, truncation_side='left', max_length=32000):
the most recent answer marker / decision survive, the oldest text is dropped.
Returns (ids, mask) when no extra masks, else (ids, mask, *extra_masks).
"""
if len(full_ids) <= max_length:
return (full_ids, mask) if not extra_masks else (full_ids, mask, *extra_masks)
cut = len(full_ids) - max_length
if not extra_masks:
return full_ids[cut:], mask[cut:]
return (full_ids[cut:], mask[cut:], *(em[cut:] for em in extra_masks))
|