mirror of
https://github.com/wassname/moral-maps.git
synced 2026-09-25 13:50:17 +08:00
wassnameandClaudypoo
9d3741fb45
administer()/read_items() now route through _rollout_natural_or_forced (the
nominal MFV core) instead of a think=0 single forward, so an activation steer
accrues over the think trace before the prefilled answer slot is read (spec
moral_aliens_engine.md, resolved decision: ordinal needs a think budget). The
only per-instrument difference is the answer-token set + the downstream reducer.
force_only on the shared core: the ordinal "(" prefill is one common char, so
natural-emission detection would match it by chance in the think trace and read
logits mid-think; surveys always force-read the answer slot. Nominal path keeps
natural emission (force_only defaults False). max_think_tokens floor is 1.
Smoke (tiny-random): ordinal + nominal both run; force-only demo reads the
forced ( slot.
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
55 lines
2.3 KiB
Python
55 lines
2.3 KiB
Python
"""tinymfv: tiny moral-foundations vignettes eval.
|
|
|
|
Forced-choice 7-way scoring on Clifford 2015 vignettes (classic) +
|
|
paraphrase configs (scifi, ai-actor). Default condition is
|
|
`other_violate` (the canonical Clifford framing); `self_violate` is
|
|
available as an opt-in ablation. Each row internally does a fwd + rev
|
|
enum-order pass for position-bias debias (inside guided_rollout).
|
|
|
|
High-level usage:
|
|
|
|
from tinymfv import evaluate
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")
|
|
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B").cuda()
|
|
rep = evaluate(model, tok, name="classic")
|
|
print(rep["table"]) # per-foundation
|
|
print(rep["top1_acc"]) # argmax accuracy vs label
|
|
print(rep["mean_js"]) # JS divergence vs label dist (in nats)
|
|
|
|
Lower-level: see `guided_rollout_forced_choice` in `tinymfv.guided`.
|
|
"""
|
|
from .data import load_vignettes, load_all_vignettes, CONFIGS, ConfigName
|
|
from .eval import evaluate, CONDITIONS
|
|
from .guided import guided_rollout_forced_choice, _DEFAULT_FORCED_FOUNDATIONS
|
|
from .instrument import Instrument, InstrItem, per_item_categorical
|
|
from .instruments import get as get_instrument, INSTRUMENTS, build_instrument
|
|
from .read import read_items, resolve_answer_ids, build_user_content
|
|
from .administer import administer
|
|
|
|
|
|
def __getattr__(name: str):
|
|
# `maps` pulls matplotlib; load it lazily so `import tinymfv` stays headless and fast for the
|
|
# numeric-only consumers (steering-lite). `tinymfv.maps.plot_*` still works -- first access
|
|
# triggers the import here.
|
|
if name == "maps":
|
|
import importlib
|
|
return importlib.import_module(".maps", __name__)
|
|
raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
|
|
|
|
|
|
# The front door. Other symbols above stay importable (e.g. read_items for item subsets,
|
|
# per_item_categorical, build_instrument) but are plumbing, kept out of `import *`.
|
|
__all__ = [
|
|
# entrypoints
|
|
"evaluate", "administer", "get_instrument", "read_items",
|
|
# types consumers build / subset
|
|
"Instrument", "InstrItem",
|
|
# data API
|
|
"load_vignettes", "load_all_vignettes", "CONFIGS", "ConfigName", "CONDITIONS",
|
|
# lower-level rollout + lazy plotting
|
|
"guided_rollout_forced_choice", "maps",
|
|
]
|
|
|