alpha-yi-f7-base-locked-20260810

Alpha yi F7 base — recovered-tokenizer lineage

This is the reviewed foundation checkpoint selected before conversational SFT. It is a 63,210,496-parameter GPT-form research model in the Alpha native format.

• format: alpha-native
• transformers_compatible: false
• selected base-parent kind: stage_b_pilot
• selected base-parent step: 5000
• selected checkpoint SHA-256: 27a3bc94e6dc2cb27fb616fe011b47e2d2b374e9499a34e4ad5f8ab4545e739d
• recovered tokenizer file SHA-256: d9222c6eeeae2d7e66d6c90518d467ea57129baabdceae9a69d447688fa01c6f
• recovered tokenizer mapping SHA-256: 25c0085f847e5304ddfc3e031f747ecd1d2b291f284ff559724bdb80d0f7ca08
• native/Hugging Face tokenizer parity probes: 1,692, with zero token-ID mismatches
• remote training base commit: ee013354f67e87ae65c84bf61b513f2b60c6aedc
• repository: ajaxdavis/alpha-yi-f7-base-locked-20260810
• reviewed checkpoint archive: ajaxdavis/alpha-yi-f7-training-artifacts-locked-20260810

This repository is the corrected lineage. The earlier ajaxdavis/alpha-yi-f7-base-20260809 repository is retained only as publicly labelled negative evidence because its stages rebuilt incompatible tokenizers. Every stage represented here used the recovered tokenizer from the valid step-1,180,000 parent, and every persisted checkpoint was mapping-verified.

The .alph file contains model parameters, AdamW optimizer state, RNG state, and tokenizer artifacts. It requires the Alpha native runtime and is not a Transformers model. A standard Hugging Face tokenizer export is included under tokenizer-recovery/hf-export-v1/, with exact token-ID parity evidence.

The selected parent was chosen after the predeclared Stage B 20,000-step diagnostic regressed and remained incoherent. Stage A, the selected Stage B pilot, and the 20,000-step diagnostic were then compared over the same 200 held-out batches and deliberately reviewed before further foundation exposure was stopped. This foundation release does not claim conversational quality; the independent visible and sealed chat gates apply to the later chat release.

Raw/licensed corpora, sealed acceptance prompts and responses, credentials, temporary transports, and superseded checkpoint payloads are excluded. Their identities, negative results, review decisions, and durable archive records are preserved in evidence here or in the versioned checkpoint archive.