inv_mse_lastpos_all

Research checkpoint from a study of tokenization (reader) invariance and its effect on robustness to adversarial re-tokenization. Fine-tuned from allenai/OLMo-2-1124-7B-Instruct.

These are research artifacts, not products. Numbers below are measured on a 200-prompt AdvBench holdout that was excluded from training by construction.

Measured

metric	value
AdvTok ASR (greedy, 200-prompt AdvBench holdout)	0.025
AdvTok ASR (t=1)	0.058
Canonical ASR (greedy, no attack)	0.02
Canonical refusal (greedy)	0.905
XSTest over-refusal (safe)	0.079
XSTest refusal (unsafe)	0.875
Alpaca over-refusal	0.03
Alpaca token F1	0.428
Alpaca NLL/token	1.274
Training cost (GPU-hours, 2xA100)	7.3
Reference (zeroshot) AdvTok ASR greedy / t=1	0.565 / 0.584
Objective	Relative residual-stream MSE to the frozen reference at the LAST PROMPT TOKEN only, averaged over ALL 32 transformer layers. No continuation is generated and nothing is teacher-forced; the loss sees one position per encoding. CVaR 0.25 over 8 MDD-sampled encodings, canonical anchored.
CAVEAT: margin_canonical	24.80 vs zeroshot 17.15. For the sibling MSE arms this inflation was shown to be the COMPLIANCE log-prob collapsing, not refusal rising; that decomposition has not been run for this arm.

AdvTok ASR is attack success rate under adversarial tokenization (Geh et al., arXiv:2503.02174), Llama-Guard-3-8B judged, greedy decoding. XSTest over-refusal is the refusal rate on safe prompts — the cost side.

Training configuration

field	value
`mode`	`prefix`
`objective`	`mse_lastpos`
`kl_direction`	`forward`
`ce_weighting`	`uniform`
`ema_beta`	`None`
`harmful_mix`	`mixed`
`harmful_fraction`	`0.286`
`num_encodings`	`8`
`cvar_quantile`	`0.25`
`max_steps`	`700`
`learning_rate`	`1e-05`
`grad_accum`	`8`
`seed`	`42`
`reference_model`	`allenai/OLMo-2-1124-7B-Instruct`
`max_new_tokens`	`128`
`prefix_tokens`	`8`
`stochastok_p`	`0.3`

Parameter drift from base (training-happened guard)

group	relative L2
attn	0.01087
embed_tokens	0.00057
lm_head	0.00000
mlp	0.01063
norm	0.00160

Caveats

Single seed. No claim of significance across seeds.
Evaluated on English AdvBench/XSTest/Alpaca only.
The safety numbers are for the specific attack studied; they do not imply robustness to other jailbreaks.