# Mossez-100M-Instruct checkpoint selection

Status: **SELECTED AS AN EXPERIMENTAL RESEARCH RELEASE CANDIDATE**

The selected checkpoint is **v4 step 216** from
`concise-sft-v4-calibration-from-v3-step555`.

This is a relative research selection, not a claim that the 100M model is a
strong, reliable, safe, or production-ready assistant. The earlier strict v4
assessment remains correct: the checkpoint fails a product-like release gate.
The project owner subsequently clarified that Mossez-100M is intended as a
compact research model and that stronger practical behavior is expected only
from later model families.

## Why v4 step 216

It has the best combined behavioral profile among the bounded v1-v5 pilots:
2/31 automatic passes, 17/31 structural passes, 13 unterminated outputs, five
repetition failures, and no generated role-marker leakage. Its validation loss
was 1.434212 and test loss 1.551534 on the v4 holdout. v5 reduced its own
teacher-forced loss but did not improve this behavioral profile.

The selected model has 100,098,048 parameters after expanding the tied
embedding matrix from 32,000 to 32,007 tokens. The full checkpoint manifest
passed. Clean FP32 and FP16 exports both reloaded with tied embeddings and
produced finite logits in real CUDA forward checks.

## Known limitations

- Weak Russian and English instruction following.
- Frequent factual, arithmetic, extraction, and exact-format errors.
- Unreliable `<|end|>`/EOS termination and residual repetition.
- Failed harmful-request refusal tests; do not rely on this model for safety.
- No production, reliability, or factual-correctness claim is made.

Selected model SHA-256: `7e1a1fc5a37287853e86751882b92a97636538fba4b6c8c066d036b281a2a1a7`  
Tokenizer v1.1 SHA-256: `e9551d84b9947f741763bf815a2d5f6bfcc47a3b67c73fcbf386223e8ed969be`