alpha_small float64 0.1 0.1 | s_vocab float64 1.38 1.38 | model_tag stringclasses 8
values | nll_clean float64 1.95 1.95 | nll_c float64 1.95 1.95 | nll_cprime float64 1.95 1.95 | nll_nulls listlengths 20 20 |
|---|---|---|---|---|---|---|
0.1 | 1.375833 | control-s17 | 1.94772 | 1.947748 | 1.947865 | [
1.9478248303332988,
1.9479425479712476,
1.9475624441262989,
1.9479425800767938,
1.9479436960935035,
1.947656661481433,
1.9474758731955946,
1.9476099649413687,
1.9475289292860367,
1.9477728300965642,
1.9478738873289676,
1.9480241960049793,
1.9478826681958987,
1.947737223370572,
1.94775491... |
0.1 | 1.375833 | control-s43 | 1.949694 | 1.949993 | 1.94952 | [
1.9500809538001478,
1.949903196557065,
1.9496386168433018,
1.9501882516528182,
1.949849650228889,
1.9499470937726648,
1.949793319093539,
1.9496145251204873,
1.9495570060799217,
1.9499853208695976,
1.9499122727950229,
1.9502315900923217,
1.9502686567831375,
1.9496388489803609,
1.949943813... |
0.1 | 1.375833 | d1-s17 | 1.946821 | 1.947005 | 1.947343 | [
1.9471289678814818,
1.9470532505797,
1.947202921192875,
1.9473759129957515,
1.947012505989164,
1.9469970521938047,
1.9472139688509689,
1.9471629691905663,
1.9467892898050347,
1.9471015870152368,
1.9472151323280513,
1.9476309605727988,
1.9473685805077297,
1.9470650076307792,
1.94732827325... |
0.1 | 1.375833 | d1-s43 | 1.949774 | 1.950006 | 1.949357 | [
1.9497466746202956,
1.949546042594195,
1.9492923437292737,
1.9500274913651603,
1.9494462715378968,
1.9495132106249449,
1.9496437482867364,
1.9496326031952886,
1.9493614418445202,
1.949557774379605,
1.9495383428466404,
1.9499764445235634,
1.9498663539071273,
1.9497685747905973,
1.94970147... |
0.1 | 1.375833 | d10-s17 | 1.946945 | 1.947283 | 1.94727 | [
1.947031548347071,
1.947213114145489,
1.9471601588385445,
1.9472779859145297,
1.947166162156947,
1.9470471873774742,
1.9472992758002716,
1.9468113424068871,
1.9468724886762454,
1.9474862057375406,
1.9471416328215767,
1.9473934405302276,
1.947583580603365,
1.9471733131788374,
1.9472169647... |
0.1 | 1.375833 | d10-s43 | 1.950147 | 1.950141 | 1.949955 | [
1.9503493871566004,
1.9500922232656903,
1.9499369542827651,
1.9503018603391893,
1.9501080866161498,
1.9496072970452856,
1.950098468911173,
1.9498844255608194,
1.94969635475994,
1.950121018171869,
1.9501000741884915,
1.9503614129171438,
1.9502314027634382,
1.9502184850270632,
1.9502565134... |
0.1 | 1.375833 | d100-s17 | 1.947412 | 1.947742 | 1.947719 | [
1.947481147857684,
1.9478850738784468,
1.9474378508762118,
1.947888097941736,
1.947595068107444,
1.947788939124248,
1.9476924087738823,
1.9474664894423384,
1.9474391740434902,
1.9475955362900639,
1.9475948084712866,
1.9479258812841822,
1.948115518696135,
1.9475340304385862,
1.94772207513... |
0.1 | 1.375833 | d100-s43 | 1.949331 | 1.949323 | 1.949216 | [
1.9493348698984543,
1.9492937024918318,
1.9494035108586385,
1.9496395509751117,
1.9494491911883656,
1.949377522675159,
1.9495360650279203,
1.9495964432767738,
1.9491018492388223,
1.9495044850353893,
1.9494002270196025,
1.949607896721056,
1.9495079207476185,
1.9491356119897383,
1.94953166... |
Fine-tune-time metadata tags — codebook, calibration, and detection references
The embedding-code codebook and the recorded detection-statistic reference values from the fine-tune-time metadata tags project. These are the small reference artifacts the code's detection and attribution results are checked against.
Contents
| Subfolder | Source training run | Files |
|---|---|---|
exp2-embedding-codes/codes/ |
Experiment 2 — embedding codes | codes.pt — the seeded code bundle: unit code c (4096-d), orthogonalized never-injected control c_prime, 8 control directions, K={4,16,64} codebooks, seeds; assignments_K{4,16,64}.csv — per-document codebook assignments |
exp5-dose-marked/detection/ |
Experiment 5 — training-time attribution | metrics.json — recorded weight-space z-scores (d100-s17: z_c = +10.684, control code z_c′ = −0.299); test3_*.json — per-run detection records |
exp8-llama-transfer/detection/ |
Experiment 8 — model-family transfer | Llama-side attribution z-scores and NLL records |
The code bundle is rebuildable bit-identically from the seeds with embedding_codes.codes in the GitHub repo.
Loading
import torch
from huggingface_hub import hf_hub_download
codes = torch.load(
hf_hub_download(
"siddharthmb/mats-gf-metadata-tags-codebook",
"exp2-embedding-codes/codes/codes.pt",
repo_type="dataset",
),
map_location="cpu", weights_only=False,
)
c = codes["c"] # unit-norm code direction, d_model = 4096
Links
- Code and reproduction anchors: https://github.com/Sid-MB/mats-gf-metadata-tags
- Paper (PDF in-repo): https://github.com/Sid-MB/mats-gf-metadata-tags/tree/main/paper
- Companion repos: adapters · probes · data
- Downloads last month
- 90