MercanAI/Mercan-0.8B-Preference-Plus-Q4

Mercan 0.8B Preference Plus — Q4_K_M

This repository contains the HH-RLHF-TR continuation-DPO Mercan 0.8B deployment artifact.

Main artifact

• Mercan-0.8B-Preference-Plus-Q4_K_M.mercan — Mercan v1 / GGUF v3, Q4_K_M
• Mercan-0.8B-Preference-Plus-Q4_K_M.mercan.sha256 — SHA-256
• Mercan-0.8B-Preference-Plus-Q4_K_M.mercan.json — conversion/provenance manifest

Training chain: continuation SFT -> DPO -> HH-RLHF-TR continuation DPO
Source checkpoint: step_00004958.pt
Base DPO checkpoint: step_00007895.pt
Preference source: MercanAI/hh-rlhf-tr

Model contract

• tokenizer: NDSRF004
• tokenizer SHA-256: 72412d981dac65a29d1767bc98821fc2bcffc2de53c534e7c719598515bfb600
• vocab rows: 32,002
• chat control IDs: 32,000 / 32,001
• canonical roles: sistem, kullanici, asistan
• hidden size: 1536
• layers: 24
• attention heads: 12
• KV heads: 4
• FFN width: 5632
• context/sliding window: 32,768
• RoPE base: 1,000,000
• MorphFFN layers: 18

A fixed 100-question TR-MMLU diagnostic sample produced 26/100, matching the pre-HH DPO checkpoint on the same frozen sample.

The structural control IDs are stored with Mercan runtime aliases <|im_start|> / <|im_end|> while canonical role headers remain Turkish.

CLI

Use the explicit artifact name: mercan run MercanAI/Mercan-0.8B-Preference-Plus-Q4:Mercan-0.8B-Preference-Plus-Q4_K_M.mercan