Request for Correction Regarding the A.X K2 Entry

This response is not intended to challenge the value of comparing and referencing architectural relationships among open-source models in this Space. Our concern is that the conclusions presented by the Space are substantially stronger and more categorical than what the publicly stated criteria and supporting evidence can justify.

In particular, labeling a model as involving "weight inheritance" is a serious factual claim, as such inheritance may directly conflict with the independence requirements of Korea's Sovereign AI Foundation Model Project. Accordingly, such a claim requires a substantially higher standard of evidence than a general observation of model similarity.

This request concerns one entry — skt/A.X-K2, which in the current build (commit 12773b9500172e91ffedcf6c5e6fc2c4931de0be, retrieved 2026-08-09) is rendered as:

1. Direct from-scratch training evidence exists

A.X K2 was developed under Korea's Sovereign AI Foundation Model Project, where independent weight initialization and full training are part of the technical independence requirement.

For A.X K2, this is not merely a statement made after release. Sequential checkpoints beginning the start of training, together with the training loss curve from the first iteration onward, were submitted through the project's formal evaluation and verification process to the Ministry of Science and ICT (MSIT), the ministry responsible for the national AI foundation model program.

This is a contemporaneous record of the model's actual initialization and training history. Inferring lineage retrospectively from final-model similarity is fundamentally weaker evidence than the training provenance itself.

2. A.X K2 is not an exact architectural match

A.X K2 and DeepSeek-V3 share hidden size, intermediate size, and layer count. They differ in every other structural field that is publicly verifiable:

The last two rows are not cosmetic. attention_output_gate, gated_norm, and gated_norm_rank are learned components active throughout pretraining — head-specific gating on the attention output, and an input-dependent gate on the normalization path. The Space's own attention table already records A.X K2's mechanism as gated MLA, distinct from the plain MLA of the config it is being matched against.

The accompanying article [2] defines its architecture fingerprint as (hidden, intermediate, layers, heads, kv) and states that what makes it a fingerprint is several fields matching simultaneously. A.X K2 matches three of those five.

These differences are also why A.X K2 registers model_type: axk2 rather than deepseek_v3. That is not a naming choice: the gated components above require their own modeling implementation, and the checkpoint cannot be loaded under DeepSeek-V3 modeling code. For contrast, kakaocorp/kanana-2-30b-a3b-instruct-2601 in the same table does register deepseek_v3 — it runs on that implementation, and that is what a match looks like under this criterion.

Calling this an "exact architectural match" is therefore not consistent with the public configs, nor with the criterion the Space itself states.

3. The tokenizers are also substantially different

A direct vocabulary-set comparison gives:

Vocabulary sizes here are taken from the vocab_size field in each model's public config.json. The shared-token count is obtained by directly comparing the published tokenizer vocabularies. Method in the appendix.

Even under the Space's own containment metric — the one that scores a strict subset as 1.0 — the overlap is about 51.78%, not the 0.60 recorded in the table, and symmetric Jaccard similarity is 29.60%. This does not support tokenizer identity or near-identity.

4. CKA should not be treated as proof of weight inheritance

CKA [1] measures how similarly two models organize a shared set of inputs in representation space. Whether one checkpoint was initialized from another is a claim about training history, not a statistic about the present model. Independently trained models can converge on similar representations through shared architecture, data, and objectives.

The accompanying article [2] states this directly: a known continued-pretraining derivative scored roughly 0.25 against 0.21 for unrelated models of similar architecture — only a small separation even where inheritance is known to have occurred — and it concludes that the weights axis is supporting evidence rather than a verdict on its own. We accept that framing; we ask only that the Space's presentation match it, rather than rendering the axis as a categorical factual state.

Requested Correction

1. Correct the architecture axis from foreign to partial. A.X K2 shares the hidden / intermediate / layer dimensions of DeepSeek-V3, but differs in head counts, dense-layer configuration, vocabulary size, and the gated attention and normalization components — which is what this Space labels elsewhere as dimensions borrowed rather than an exact match.
2. Correct the weights axis from inherited to from-scratch, absent direct and reproducible evidence of checkpoint inheritance.
3. Correct the resulting genotype from Ported to Adapted. This follows mechanically from (1) and (2) under the Space's own genotype rule (architecture × weights); no separate judgment is required.
4. Distinguish CKA-based representational similarity from checkpoint inheritance history in the table and tooltips. The former is a similarity measure; the latter is a factual claim about how a model was initialized and trained.
5. Publish the tokenizer measurement (revision, normalization) behind the 0.60 figure so the difference from the values above can be reconciled.

Corrections (1) and (2) are independent of each other, and each on its own already moves the entry out of the Ported category under the same rule.

We welcome technical criticism and reproducible analysis. Our request is only that the strength of the label match the strength of the evidence behind it.

---

Appendix — tokenizer measurement

Vocabulary sets were taken from the published tokenizer.json of each model and compared directly as sets of token strings. Containment is |A ∩ B| / min(|A|, |B|), the metric defined in the accompanying article; Jaccard is |A ∩ B| / |A ∪ B|. Revisions used: A.X K2 7b3969c2a84b43c03718b9acc82e15a68b1a4c5a, DeepSeek-V3 e815299b0bcbac849fa540c768ef21845365c9eb.

References

[1] Kornblith, Simon, et al. "Similarity of neural network representations revisited." International Conference on Machine Learning, PMLR, 2019.

[2] https://huggingface.co/blog/mayafree/model-dna