Evidence Map

This page maps each public claim to the supporting script family and sanitized evidence snippet.

Claim: Tiny TQ2 training progressed far enough to save checkpoint step 6

Date: 2026-03-23
Script: run_finetune_tiny.sh
Evidence: 2026-03-23-finetune-tiny-snippet.txt
Expected supporting lines: Training for 1 epochs, Optimizer: adamw, loss=3.04376, checkpoint_step_00000006/model.gguf

Claim: Fast TQ2 checkpoint inference produced a non-empty completion

Date: 2026-03-23
Script: run_infer_tq2_checkpoint6_fast.sh
Evidence: 2026-03-23-infer-tq2-checkpoint6-fast-snippet.txt
Expected supporting lines: Mali-G52 MC2, offloaded 25/25 layers to GPU, checkpoint_step_00000006/model.gguf, A:, total time =

Claim: TQ1 published-adapter rerun provides a single-run throughput reference

Date: 2026-03-25
Command family: direct llama-cli rerun over the existing Windows -> adb forward -> SSH -> Termux path
Evidence: 2026-03-25-infer-tq1-snippet.txt
Expected supporting lines: $HOME/qvac-bitnet/models/bitnet-xl.tq1_0.gguf, $HOME/qvac-bitnet/adapters/tq1_0-biomed-trained-adapter.gguf, prompt eval time =, eval time =, total time =

Claim: TQ2 fast checkpoint rerun provides a shortened single-run throughput reference

Date: 2026-03-25
Command family: direct llama-cli rerun over the existing Windows -> adb forward -> SSH -> Termux path
Evidence: 2026-03-25-infer-tq2-fast-snippet.txt
Expected supporting lines: $HOME/qvac-bitnet/models/bitnet-xl.tq2_0.gguf, $HOME/qvac-bitnet/checkpoints-tiny/checkpoint_step_00000006/model.gguf, prompt eval time =, eval time =, total time =