chakibmed/whispbook-functiongemma-270m-speaker-attribution-mlx-lora-hpo-gen4-qv-step575

Whispbook FunctionGemma Speaker Attribution MLX LoRA HPO Gen4 QV Step 575

This repository contains an MLX-LM LoRA adapter for FunctionGemma speaker attribution.

It is the current local HPO leader for the not_in_candidates phase. The adapter is tuned to detect when the speaker is absent from the candidate list. It is still a research checkpoint: it improves fallback discovery, but it can falsely reject known speakers.

Base Model

• mlx-community/functiongemma-270m-it-4bit

Selected Trial

• Trial: gen4-r0075-qv-r8-d010-lr5e6-b6-s575-save25
• Checkpoint: 0000575_adapters.safetensors
• Target modules: self_attn.q_proj, self_attn.v_proj
• LoRA rank: 8
• LoRA dropout: 0.10
• Learning rate: 5e-6
• Batch size: 6
• Iterations: 575
• not_in_candidates training ratio: 0.075

HPO Probe Result

On the 100-row HPO probe:

metric	value
HPO score	0.0078
total accuracy	35 / 100
known-speaker accuracy	5 / 58
not_in_candidates accuracy	30 / 42
false fallback on known speakers	9
forced known-speaker pick on fallback examples	11

Larger Local Check

On a 500-row local generation check using 458 known-speaker rows and 42 not_in_candidates rows:

metric	value
HPO score	0.0190
total accuracy	68 / 500
known-speaker accuracy	38 / 458
not_in_candidates accuracy	30 / 42
false fallback on known speakers	67
forced known-speaker pick on fallback examples	11

For comparison:

adapter	HPO score	known	not_in_candidates	false fallback	forced pick
Gen4 QV step 575	0.0190	38 / 458	30 / 42	67	11
Gen3 QV step 575	-0.0649	46 / 458	27 / 42	83	14
Conservative QVO step 500	-0.1050	51 / 458	6 / 42	4	36

This means Gen4 is currently the strongest fallback detector, while the QVO checkpoint remains the safer choice when false fallback must be minimized.

Files

• adapters.safetensors: selected MLX LoRA adapter weights.
• adapter_config.json: MLX-LM adapter metadata.
• trial_config.json: HPO trial configuration.
• eval_0000575_summary.json: 100-row generation probe summary.
• eval_0000575_full500_summary.json: larger 500-row generation-check summary.
• hpo_leaderboard.json: leaderboard snapshot after this phase.
• functiongemma_lora_config.yaml: MLX-LM training configuration used for this trial.
• report/training_report.html: Plotly checkpoint evolution report.
• report/checkpoint_metrics.json: checkpoint metrics used by the report.
• report/README_metrics.md: markdown summary for readme/documentation work.