MercanAI/Mercan-0.8B-SFT-step19838

Mercan 0.8B SFT — final step 19838

Final full-parameter Turkish SFT checkpoint exported to Mercan v1.

Artifact

• —File: model.mercan
• —Physical container: GGUF v3
• —Weight format: F16 matrices, F32 1-D tensors
• —Parameters: ~823M
• —Vocabulary: 32,002 model IDs
• —Context length: 32,768
• —Sliding window: 32,768
• —Layers: 24
• —Hidden size: 1,536
• —FFN size: 5,632
• —Attention heads / KV heads: 12 / 4
• —RoPE theta: 1,000,000
• —MorphFFN layers: 18
• —SHA-256: edc7443c104047c7fac8bae7131472bc7069b66557ce1d3021d7a167968d9abb

Tokenizer and chat contract

The base tokenizer is NDSRF004. The original 32,000-token surface vocabulary is preserved. The final SFT checkpoint adds two structural chat IDs:

• —32000 — message start
• —32001 — message end

For Mercan container interoperability these are exposed as the control-token aliases <|imstart|> and <|imend|>.

The model was trained with Turkish role headers:

• —sistem
• —kullanici
• —asistan

Canonical serialization:

MESSAGESTART + encode(role + newline) + encode(content) + MESSAGEEND + encode(newline)

For assistant generation, append MESSAGE_START followed by encode(asistan + newline).

Runtime compatibility note

The current refactor Mercan CLI wrapper formats roles as user / assistant. This checkpoint was trained with kullanici / asistan role text, so callers should use the Turkish role contract above, or update the wrapper, to match SFT training exactly.

Provenance

Source checkpoint: step_00019838.pt

This export follows PTTOMERCAN.md, with architecture overrides matching the trained 32,002-vocabulary and 32,768-context model.