text stringlengths 0 151 |
|---|
========================================================================================== |
LFM2.5-8B-A1B ROUTER LOGIT + EXPERT L2 NORM ANALYSIS |
========================================================================================== |
Generated: 2026-07-21T22:30:41.152699Z |
Model: LiquidAI/LFM2.5-8B-A1B |
Input: /kaggle/input/datasets/wahmaddeepthink/perfecttext/perfectblend_assistant_texts.txt |
Tokens: 1000 |
MoE layers: 22 |
Experts/layer: 32 |
Trained top-k: 4 |
========================================================================================== |
METRIC EXPLANATIONS |
========================================================================================== |
LOGIT METRICS (raw router logits, before any sigmoid / top-k / normalization) |
------------------------------------------------------------------------------ |
mean_top1 : Mean of the highest raw logit per token. |
std_top1 : Std of highest raw logits across tokens. High = router is |
selectively confident (some tokens easy, some hard). |
mean_topk : Mean of the average of top-4 raw logits per token. |
std_topk : Std of that average across tokens. |
mean_gap_top1_to_topk : Mean of (top1_logit - mean(top4_logits)) per token. |
Large gap = top-1 dominates; small = top-4 are close. |
std_gap_top1_to_topk : How much that gap varies across tokens. |
mean_std_per_token : Mean of std(all 32 logits) per token. High = the router |
strongly separates winners from losers within a token. |
mean_range_per_token: Mean of (max - min) raw logit per token. |
L2 NORM METRICS (output magnitude of each expert, ALL 32 evaluated per token) |
------------------------------------------------------------------------------ |
pct_top1_logit_is_max_l2 : % of tokens where the expert with highest logit |
also produces the largest L2 output. 100% = perfect |
agreement between routing score and output magnitude. |
mean_rank_l2_of_top1 : Mean L2-rank of the top-1 logit expert (1 = highest L2, |
32 = lowest). Low = the router's #1 pick is also the |
biggest producer. |
mean_l2_top1_expert: Mean L2 norm of the top-1 logit expert's output. |
mean_l2_topk_experts: Mean L2 norm of the top-4 logit experts' outputs. |
mean_l2_all_experts: Mean L2 norm across ALL 32 experts. |
mean_max_l2 : Mean of the maximum L2 norm per token. |
mean_ratio_top1_to_mean_l2 : Mean of (l2_top1 / mean_l2_all). >1 means the |
router's pick is larger than average. |
mean_ratio_top1_to_max_l2 : Mean of (l2_top1 / max_l2). Close to 1.0 means |
the router's pick is almost always the biggest producer. |
cv_l2_mean : Coefficient of variation = std(L2s)/mean(L2s), averaged |
over tokens. High = huge spread between experts. |
cv_l2_max : CV using max as denominator = std(L2s)/max(L2), averaged. |
This is insensitive to layer-wise magnitude scaling. |
High = even the best expert is not that far ahead of the |
pack; low = one expert towers over the rest. |
corr_top1_logit_vs_top1_l2 : Pearson correlation across tokens between the |
top-1 logit and the L2 norm of that same expert. |
corr_top1_logit_vs_max_l2 : Pearson correlation across tokens between the |
top-1 logit and the maximum L2 norm (any expert). |
corr_topk_logit_vs_topk_l2_by_rank : For each rank r=1..4, correlation across |
tokens between the logit at rank r and the L2 norm of |
the expert occupying that rank. |
mean_corr_topk_logit_l2_per_token : For each token, correlate the 4 selected |
logits with their 4 L2 norms; average across tokens. |
Tells us: within the chosen set, do higher scores mean |
bigger outputs? |
========================================================================================== |
LAYER 2 |
========================================================================================== |
Logits | top1=+0.8838 (std=0.4793) | topk_mean=-0.0492 (std=0.3317) |
| gap_top1_to_topk=+0.9330 (std=0.3456) | per_token_std=0.8079 range=3.7588 |
L2 | top1_is_max_l2=53.2% | rank_l2_of_top1=3.25 |
| l2_top1=0.2889 | l2_topk=0.2515 | l2_all=0.1861 | max_l2=0.3227 |
| ratio_top1/mean=1.5669 | ratio_top1/max=0.9021 |
| CV(mean)=0.2339 | CV(max)=0.1342 |
Corr | top1_logit vs top1_l2=+0.1191 | top1_logit vs max_l2=+0.1444 |
| per-token topk_logit~l2 corr=+0.5545 |
| by-rank corr=[0.1191, 0.2397, 0.2071, 0.2534] |
Samples (first 5 tokens): |
[ 0] 'We' top1_logit=-0.6133 E21 topk_logits=[-0.6133, -1.2656, -1.2969, -1.5391] l2_top1=0.41 max_l2=0.42(E25) rank_l2_of_top1=2 |
[ 1] ' are' top1_logit=+0.0267 E29 topk_logits=[0.0267, -0.4922, -0.5781, -0.6562] l2_top1=0.35 max_l2=0.44(E17) rank_l2_of_top1=6 |
[ 2] ' given' top1_logit=+0.2930 E29 topk_logits=[0.293, 0.1689, 0.0234, -1.2109] l2_top1=0.33 max_l2=0.38(E11) rank_l2_of_top1=6 |
[ 3] ' that' top1_logit=+0.0532 E22 topk_logits=[0.0532, -0.582, -0.8555, -0.8828] l2_top1=0.50 max_l2=0.50(E22) rank_l2_of_top1=1 |
[ 4] ' the' top1_logit=+0.3672 E 2 topk_logits=[0.3672, 0.2393, -0.9414, -1.1719] l2_top1=0.35 max_l2=0.45(E17) rank_l2_of_top1=6 |
========================================================================================== |
LAYER 3 |
========================================================================================== |
Logits | top1=+0.5271 (std=0.4765) | topk_mean=-0.4637 (std=0.2819) |
| gap_top1_to_topk=+0.9908 (std=0.3951) | per_token_std=0.7699 range=3.6176 |
L2 | top1_is_max_l2=56.1% | rank_l2_of_top1=2.15 |
| l2_top1=0.2799 | l2_topk=0.2411 | l2_all=0.1916 | max_l2=0.3024 |
| ratio_top1/mean=1.4660 | ratio_top1/max=0.9284 |
| CV(mean)=0.1696 | CV(max)=0.1062 |
Corr | top1_logit vs top1_l2=+0.3636 | top1_logit vs max_l2=+0.3086 |
| per-token topk_logit~l2 corr=+0.6607 |
| by-rank corr=[0.3636, 0.2691, 0.1928, 0.0767] |
Samples (first 5 tokens): |
[ 0] 'We' top1_logit=-0.2754 E 5 topk_logits=[-0.2754, -0.7812, -1.1328, -1.2109] l2_top1=0.27 max_l2=0.30(E23) rank_l2_of_top1=5 |
[ 1] ' are' top1_logit=+0.7695 E29 topk_logits=[0.7695, 0.5859, -1.5547, -1.7422] l2_top1=0.28 max_l2=0.28(E29) rank_l2_of_top1=1 |
[ 2] ' given' top1_logit=-0.1494 E26 topk_logits=[-0.1494, -0.3223, -0.3867, -1.125] l2_top1=0.27 max_l2=0.38(E21) rank_l2_of_top1=9 |
End of preview. Expand in Data Studio
No dataset card yet
- Downloads last month
- -