direct_gen_subschema

Checkpoints and CypherBench inference outputs of the direct sub-schema Text-to-Cypher experiment (branch direct_gen_subschema of fisherman611/cypher-extract). One multitask model does two tasks: question + full schema -> sub-schema, then question + sub-schema -> Cypher.

• Data: fisherman611/cypher-extract-direct-subschema-data (llamafactory_direct_subschema_v1, 13,654 train rows = 6,827 questions x 2 tasks).
• Training: 3 epochs (427 steps/epoch, 1281 steps), batch 2 x 8 accumulation x 2 GPUs, cutoff_len 2048. Final checkpoint is checkpoint-1281. Teacher: Qwen3-4B-Instruct-2507 LoRA (r=32, alpha=64). Students: Qwen3-0.6B (sft full fine-tune; csd, distillm_adaptive_srkl, amid, hpd LoRA KD from the teacher's checkpoint-1281).
• Checkpoint directories keep the DeepSpeed training state (optimizer, global_step*), so sft checkpoints are large.

Layout

• direct_subschema/lora/qwen3/<method>/checkpoint-* for teacher_lora, csd, distillm_adaptive_srkl, amid, hpd.
• direct_subschema/full_finetune/qwen3/sft/checkpoint-* (the last two checkpoints only).
• inference/direct_subschema/<setting>/qwen3/seed42/<method>/cypherbench/ with subschema_predictions.jsonl, generator_predictions.jsonl, metrics.json (sub-schema quality) and run_config.json.

CypherBench test, seed 42, final checkpoint

model	EX	PSJS	executable
teacher_lora	81.22	78.09	99.06
sft	57.79	52.09	98.30
hpd	56.09	49.92	98.04
distillm_adaptive_srkl	56.05	49.94	98.64
amid	54.09	47.98	97.23
csd	53.71	47.24	97.53

One seed and one training run per model. Details and sub-schema metrics: docs/benchmark-results/direct-subschema-cypherbench-seed42.md.