banhchungtuongot/qwen25_1p5b_memory_fix_v10g_b_unified_adapter

Model Card for sftmemoryfixv10gb_unified

LoRA SFT adapter (rank=32, alpha=64, RSLoRA, all-linear, bf16) — v10g-b, the first model to pass all 3 targets simultaneously.

Training mix: 2705 samples (base v9 2441 + abstentionpatchv1×3 + v2tp×3 + writepatchv10g×1). Merged result: `banhchungtuongot/qwen251p5bmemoryfixv10gb_unified`.

Training procedure

This model was trained with SFT on a memory-aware conversational task. The assistant learns to:

• —Write stable personal facts to memory (memory_ops: [{op: upsert, key, value}])
• —Abstain from writing transient state, third-party info, or retrieval prompts
• —Return structured JSON: {response, memory_ops, needs_retrieval, fallback}

Framework versions

• —PEFT 0.17.1
• —TRL: 1.4.0
• —Transformers: 4.57.2
• —Pytorch: 2.6.0+cu124