mirror of
https://github.com/wassname/evil_MoE.git
synced 2026-08-11 03:40:06 +08:00
wassnameandClaudypoo
b53043cec3
Cleanup by a prior agent, verified green here: 'just smoke' (erase arm) runs end-to-end and all four wired gates pass (verify_rewards 52/52, verify_eval_gap, verify_partition, verify_science_invariants). - train.py -318 lines: Config dataclass -> train_config.py, checkpoint/ deploy-artifact IO -> run_artifacts.py. - results.py / results_deploy.py / probe_distill.py slimmed. - drop stale derived csvs under out/figs (a5_generalisation, dyn_*, substrate_aggregate, train_vs_deploy_60). - gitignore /.pi/ panel scratch. Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
115 lines
3.1 KiB
Python
115 lines
3.1 KiB
Python
"""Typed CLI configuration for train.py."""
|
|
from __future__ import annotations
|
|
|
|
from dataclasses import dataclass
|
|
from pathlib import Path
|
|
from typing import Literal
|
|
|
|
from .rewards import EnvMode
|
|
|
|
|
|
@dataclass(kw_only=True)
|
|
class Config:
|
|
intervention: Literal["none", "erase", "routeV"] = "erase"
|
|
adapter: Literal["antipasto", "lora_frozen_b"] = "antipasto"
|
|
lora_r: int = 32
|
|
lora_b_seed: int = 0
|
|
|
|
model: str = "Qwen/Qwen3-4B"
|
|
steps: int = 100
|
|
group: int = 6
|
|
max_new: int = 1024
|
|
n_problems: int = 992
|
|
beta: float = 0.0
|
|
prompts_per_step: int = 8
|
|
lr: float = 7e-5
|
|
adam_beta1: float = 0.9
|
|
adam_beta2: float = 0.99
|
|
clip: float = 0.2
|
|
weight_decay: float = 0.1
|
|
warmup_frac: float = 0.1
|
|
grad_clip: float = 10.0
|
|
seed: int = 41
|
|
unbiased: bool = True
|
|
|
|
preserve_magnitude: bool = True
|
|
gate_mode: Literal["one_sided", "no_gate", "reverse"] = "one_sided"
|
|
project_overshoot: float = 1.0
|
|
v_hack_path: Path | None = None
|
|
v_hack_extract_top_k: int = 12
|
|
v_hack_k: int = 5
|
|
v_hack_tau_axis: float = 0.0
|
|
v_hack_drop_bottom_frac: float = 0.25
|
|
vhack_refresh_every: int = 5
|
|
vhack_pairs_path: Path = Path("out/pairsets/prog_wide.json")
|
|
|
|
routeV_random_v_seed: int | None = None
|
|
routeV_per_token: bool = False
|
|
routeV_gate: Literal["grad_cosine", "act_vote", "online_stats"] = "grad_cosine"
|
|
routeV_absorb_all: bool = False
|
|
online_stats_lo: float = 0.05
|
|
online_stats_hi: float = 0.95
|
|
rollout_ablate_frac: float = 0.0
|
|
|
|
env_mode: EnvMode = "run_tests"
|
|
unhackable_frac: float = 0.0
|
|
teacher_pool_dir: Path | None = None
|
|
mix_ratio: float = 0.125
|
|
teacher_off_step: int | None = 30
|
|
teacher_modes: tuple[str, ...] | None = None
|
|
|
|
eval_ablate_every: int = 0
|
|
eval_n_prompts: int = 32
|
|
eval_batch_size: int = 2
|
|
save_ckpt_every: int = 10
|
|
cos_pre_split_every: int = 1
|
|
half_a: str = ""
|
|
out_tag: str = ""
|
|
|
|
@property
|
|
def preset_name(self) -> str:
|
|
return type(self).__name__.removesuffix("Config").lower() or "base"
|
|
|
|
@property
|
|
def arm(self) -> str:
|
|
return {"none": "vanilla", "erase": "projected", "routeV": "routingV"}[self.intervention]
|
|
|
|
|
|
@dataclass(kw_only=True)
|
|
class SmokeConfig(Config):
|
|
model: str = "llamafactory/tiny-random-qwen3"
|
|
steps: int = 30
|
|
group: int = 4
|
|
max_new: int = 32
|
|
n_problems: int = 100
|
|
beta: float = 0.0
|
|
prompts_per_step: int = 1
|
|
|
|
|
|
@dataclass(kw_only=True)
|
|
class FastConfig(Config):
|
|
model: str = "Qwen/Qwen3-4B"
|
|
steps: int = 60
|
|
teacher_pool_dir: Path | None = Path("out/pools/teacher_pool_runtests_dense")
|
|
vhack_pairs_path: Path = Path("out/pairsets/prog_wide.json")
|
|
grad_clip: float = 500.0
|
|
group: int = 8
|
|
max_new: int = 512
|
|
n_problems: int = 200
|
|
beta: float = 0.0
|
|
prompts_per_step: int = 4
|
|
lr: float = 3e-3
|
|
adam_beta1: float = 0.5
|
|
adam_beta2: float = 0.9
|
|
|
|
|
|
@dataclass(kw_only=True)
|
|
class FullConfig(Config):
|
|
model: str = "Qwen/Qwen3-4B"
|
|
steps: int = 200
|
|
group: int = 4
|
|
max_new: int = 1536
|
|
n_problems: int = 992
|
|
beta: float = 1e-3
|
|
prompts_per_step: int = 64
|