creg-sql-mlx-community-qwen2-5-coder-3b-instruct-4bit-mlx-4bit

hbmartin/creg-sql-mlx-community-qwen2-5-coder-3b-instruct-4bit-mlx-4bit

An MLX 4-bit text-to-SQL derivative for the frozen synthetic CREG commercial real-estate portfolio. This artifact is a research prototype, not a general SQL model.

Reproducibility

• Base: mlx-community/Qwen2.5-Coder-3B-Instruct-4bit@3dd939c621c08e5753d5b89f35a2642cd83b98ca
• Verified base artifact tree SHA-256: bb07390155226383c705229516cb9e333fcbf58e38d490a768bac78da530cd4b
• Base artifact lock SHA-256: 7dbb948f5b6cf7d9e773cecb2de15195ca542532337cad75a7b0e9161cc70974
• Code revision: a35f6c06ad031e91512ba3977c3430d241b59493 (dirty state: true)
• Training run: qlora-qwen25-coder-3b-seed-424242
• Training runner SHA-256: b2765239ade40bd3231f057de1aec431ac227e7a8ce9769fc4651df29303943c
• Corpus generator SHA-256: 416f1370045de5fffe69a557fa1486fd80f4de645a7f4311b9ae878e3ccd79c9
• Model manifest input SHA-256: 3c5c293da1760ea054862d35ddd2e231f458a5bfec8d1386e5fd4bb72044fb9c
• Pinned Python lock SHA-256: 034b816d176c52f6a0ba5eb5f657f1dc1a7fbd222edd564e32e75361adb91cd3
• Training configuration SHA-256: 63a139bb79e5ad77d0241b9b76d0c2892a3459019bcfad04367d8d932800b545
• Corpus manifest SHA-256: 4e3d25923a8426ba3c09349d240a1c77aca79daf5c69631694d28dfe4374cabe
• Gold set remained held out: 2bde4dedc23bc7938d0250f2e2d5e22502903c8caad6c0951e1bbca9d0d77036
• Adapter tree SHA-256: b31b6b1997091f796d83109bf3bba91856a7c301eedf3c78541d949fcd3ae349
• Training log SHA-256: bd81f135e0699d846a4a08ca1552c74c48a6fefca6b419e1d727a3d90e59ae32
• Fused output tree SHA-256 before publication documentation: 5cbda9c9115f01df2a8a3a23a0de0a07447de6f3a5053f8c05a89c6bf33a3aa6
• Model payload SHA-256 excluding documentation, license, and notice files: 518a3ab2d819f397250d681f96a5b9d8fd81ae441c770b8fc557d294217d85b9
• Quantization: fused 4-bit affine, group size 64
• Commercial use allowed by the declared inherited license: false

Training corpus inputs:

• fine-tuning/synth/out/train.jsonl: 3a9ad4806692cdc89e8e68c77e29c5e1eedaefac5745c3a87bd4e4fb1758021e (byte-for-byte regeneration: true)
• fine-tuning/synth/out/valid.jsonl: b0e72fde78f50e80bc5bdd4664eb7e87695db8a52e042067cfb32ddfc0fcec33 (byte-for-byte regeneration: true)
• fine-tuning/synth/out/gate_stats.json: d63b2ae38ec22276dc7706c34d9f12b5f301909e50f01c3c7412043abd22f442 (byte-for-byte regeneration: true)

The complete YAML configuration uses seed 424242, 600 iterations, batch size 4, 16 adapted layers, learning rate 1e-4, prompt masking, and explicit mlx-lm defaults. The immutable training run retains the complete commands, per-file adapter inventory, training log, and fused output inventory.

# Complete replacement for the incompletely recorded PR #1 training command.
# The finalist runner must override `model` and `adapter_path`; sentinel values
# make a missing override fail rather than selecting an implicit upstream.
model: REQUIRED_FINALIST_MODEL_OVERRIDE
train: true
fine_tune_type: lora
optimizer: adam
optimizer_config:
  adam: {}
  adamw: {}
  muon: {}
  sgd: {}
  adafactor: {}
data: synth/out
seed: 424242
num_layers: 16
batch_size: 4
iters: 600
val_batches: 25
learning_rate: 0.0001
steps_per_report: 10
steps_per_eval: 200
grad_accumulation_steps: 1
resume_adapter_file: null
adapter_path: REQUIRED_IMMUTABLE_ADAPTER_PATH_OVERRIDE
save_every: 100
test: false
test_batches: 500
max_seq_length: 2048
config: null
grad_checkpoint: false
clear_cache_threshold: 0
lr_schedule: null
lora_parameters:
  rank: 8
  dropout: 0.0
  scale: 20.0
mask_prompt: true
report_to: null
project_name: creg-sql

Evaluation

• gold_v2.jsonl; GCD on; temperature 0.0; seed 0: EX 0.525, valid SQL 0.895, p95 2972846 μs; immutable run matrix-fine-tune-gold-v2-ft-qwen25-coder-3b-gcd-on-t-0_0-s-0 (manifest SHA-256 d4fc59848b564687150548b3fb18c41e1264d7a3288f5a20c2b9e781a4de3b93, summary SHA-256 0b36868b348f7f22aac919f726cd562eaa1ad292277f04c92d0c5de0fb879ec6)
• gold_v2.jsonl; GCD off; temperature 0.0; seed 0: EX 0.525, valid SQL 0.860, p95 1559067 μs; immutable run matrix-fine-tune-gold-v2-ft-qwen25-coder-3b-gcd-off-t-0_0-s-0 (manifest SHA-256 81c926f40d9c922cdcd5f73a9ceee3928ab76d610512e76ca0ffc12c2a4ac645, summary SHA-256 37c09bfc5e42df537c82268c7a7614e7488313ed525a9781b1a6e63f62790f87)

Execution accuracy is order-insensitive typed row-multiset equality with four-decimal half-even numeric normalization. These scores apply only to the frozen CREG schema/database/gold set and their immutable run manifests.

Limitations

• Narrow synthetic domain and fixed SQLite schema.
• May generate semantically incorrect, incomplete, or non-executable SQL.
• Not evaluated for arbitrary databases, adversarial prompts, or production financial decision-making.
• Generated SQL must execute under a read-only connection and should be independently reviewed.

License and required notice

This is a modified derivative of Qwen2.5-Coder-3B-Instruct. It is provided under the Qwen Research License included in this repository, together with any additional upstream license file identified by the base artifact. Non-commercial use only. Built/Improved using Qwen. Qwen, the base-model authors, and any intermediate model authors are attributed through the base-model link, NOTICE, modification notice, and included license files.