llama-finance-per-head-top-t-256-key-value-num-tokens-512

Configuration

• Base model: meta-llama/Llama-3.2-3B-Instruct
• Phase: continual
• Granularity: per-head
• top_t: 256
• Key/Value training: key + value (keys trainable)
• Cartridge tokens: 512
• Epochs: 1
• Last checkpoint: cache-step234.pt

Files

• cache-step234.pt — trained cartridge KV cache (load via KVFromLocal).
• bg_stats.pt — background token statistics used for sparse TF-IDF gating.
• config.yaml — full training configuration.
• sparse / TF-IDF ranking logs where present.