kent/consciousness
1
0
Fork 1
Code Issues 1 Pull requests 5 Projects Releases Packages Wiki Activity Actions

amygdala: underscore-prefixed files join every concept's negative pool

Browse source 
Files in direct/ named _*.txt (e.g. _baseline.txt) are conceptless
neutral prose — they should not appear as positive training signal,
but are useful as shared negatives across every concept.

Previously _*.txt files were silently skipped. Now:
  * they're loaded like any other description file;
  * concepts (the positive label set) filters them out;
  * their descriptions are concatenated into neg_pool_extra and
    extended onto every concept's neg_pool alongside the cross-concept
    negatives.

A concept's negative pool is thus "other concepts' descriptions +
everything from _*.txt files". The extra pool is announced at startup
so the user can see how many neutral samples are active.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
This commit is contained in:
Kent Overstreet 2026-04-24 11:53:42 -04:00
parent 5908b837e8
commit 6fedc9b2a8
1 changed files with 17 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

23
training/amygdala_training/train_direct.py
View file

@ -35,12 +35,11 @@ from steering_vectors.aggregators import pca_aggregator
def _load_descriptions(direct_dir: Path) -> dict[str, list[str]]: def _load_descriptions(direct_dir: Path) -> dict[str, list[str]]:
"""Each file in direct_dir is `{concept}.txt`. Descriptions are """Each file in direct_dir is `{concept}.txt`. Descriptions are
separated by blank lines within the file.""" separated by blank lines within the file. Files starting with `_`
are not concepts but are included in negative pools (e.g. _baseline.txt)."""
out: dict[str, list[str]] = {} out: dict[str, list[str]] = {}
for f in sorted(direct_dir.glob("*.txt")): for f in sorted(direct_dir.glob("*.txt")):
if f.name.startswith("_"): concept = f.stem # underscore-prefixed names keep their prefix
continue
concept = f.stem
text = f.read_text() text = f.read_text()
descs = [d.strip() for d in text.split("\n\n") if d.strip()] descs = [d.strip() for d in text.split("\n\n") if d.strip()]
out[concept] = descs out[concept] = descs
@ -69,11 +68,19 @@ def main() -> None:
target_layers = [int(x) for x in args.target_layers.split(",")] target_layers = [int(x) for x in args.target_layers.split(",")]
dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[args.dtype] dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[args.dtype]
descriptions = _load_descriptions(Path(args.direct_dir)) all_descriptions = _load_descriptions(Path(args.direct_dir))
concepts = sorted(descriptions.keys()) # Files starting with `_` are neg-pool helpers (e.g. _baseline.txt), not concepts.
concepts = sorted(k for k in all_descriptions if not k.startswith("_"))
neg_pool_extra: list[str] = []
for k, ds in all_descriptions.items():
if k.startswith("_"):
neg_pool_extra.extend(ds)
descriptions = {k: all_descriptions[k] for k in concepts}
print(f"Loaded {len(concepts)} concepts with direct descriptions:") print(f"Loaded {len(concepts)} concepts with direct descriptions:")
for c in concepts: for c in concepts:
print(f" {c}: {len(descriptions[c])} descriptions") print(f" {c}: {len(descriptions[c])} descriptions")
if neg_pool_extra:
print(f"Plus {len(neg_pool_extra)} neutral/baseline descriptions added to every concept's negative pool")
print(f"\nLoading {args.model} ({args.dtype}) on {args.device}...") print(f"\nLoading {args.model} ({args.dtype}) on {args.device}...")
tokenizer = AutoTokenizer.from_pretrained(args.model) tokenizer = AutoTokenizer.from_pretrained(args.model)
@ -117,6 +124,10 @@ def main() -> None:
for other, other_descs in descriptions.items(): for other, other_descs in descriptions.items():
if other != concept: if other != concept:
neg_pool.extend(other_descs) neg_pool.extend(other_descs)
# Underscore-prefixed files (e.g. _baseline.txt) contribute to
# every concept's negative pool, independent of the other-
# concept negatives.
neg_pool.extend(neg_pool_extra)
rng = random.Random(hash(concept) & 0xFFFFFFFF) rng = random.Random(hash(concept) & 0xFFFFFFFF)
samples: list[SteeringVectorTrainingSample] = [] samples: list[SteeringVectorTrainingSample] = []