jastorj/snowflake_arctic_text2sql_r1_7b-nl2sqlpp-16bit-v5.7.8_phase_3-cw-29K

Snowflake/Arctic-Text2SQL-R1-7B Fine-tuned for NL2SQL++ v8

This model is a fine-tuned version of Snowflake/Arctic-Text2SQL-R1-7B on the NL2SQL++ v8 dataset with code-with-thought reasoning.

Model Details

• Base Model: Snowflake/Arctic-Text2SQL-R1-7B
• Task: Text-to-SQL generation
• Dataset: NL2SQL++ v8 with code-with-thought reasoning
• Fine-tuning Method: LoRA (Low-Rank Adaptation) with Unsloth
• Quantization: 16-bit merged weights
• Training Dataset Size: (10055, 1) examples
• Validation Dataset Size: (1062, 1) examples

Training Configuration

• output_dir: trainer_output
• overwrite_output_dir: False
• do_train: False
• do_eval: True
• do_predict: False
• eval_strategy: IntervalStrategy.STEPS
• prediction_loss_only: False
• per_device_train_batch_size: 8
• per_device_eval_batch_size: 8
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 16
• eval_accumulation_steps: 10
• eval_delay: 0
• torch_empty_cache_steps: None
• learning_rate: 1e-05
• weight_decay: 0.01
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1.0
• num_train_epochs: 3.0
• max_steps: -1
• lr_scheduler_type: SchedulerType.COSINE
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.1
• warmup_steps: 0
• log_level: passive
• log_level_replica: warning
• log_on_each_node: True
• logging_dir: trainer_output/runs/May22_13-38-25_ip-172-31-10-229.ap-northeast-1.compute.internal
• logging_strategy: IntervalStrategy.STEPS
• logging_first_step: False
• logging_steps: 3
• logging_nan_inf_filter: True
• save_strategy: SaveStrategy.BEST
• save_steps: 50
• save_total_limit: 2
• save_safetensors: True
• save_on_each_node: False
• save_only_model: False
• restore_callback_states_from_checkpoint: False
• no_cuda: False
• use_cpu: False
• use_mps_device: False
• seed: 3407
• data_seed: None
• jit_mode_eval: False
• bf16: True
• fp16: False
• fp16_opt_level: O1
• half_precision_backend: auto
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• local_rank: 0
• ddp_backend: None
• tpu_num_cores: None
• tpu_metrics_debug: False
• debug: []
• dataloader_drop_last: False
• eval_steps: 50
• dataloader_num_workers: 0
• dataloader_prefetch_factor: None
• past_index: -1
• run_name: None
• disable_tqdm: False
• remove_unused_columns: True
• label_names: None
• load_best_model_at_end: True
• metric_for_best_model: eval_loss
• greater_is_better: False
• ignore_data_skip: False
• fsdp: []
• fsdp_min_num_params: 0
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• fsdp_transformer_layer_cls_to_wrap: None
• accelerator_config: AcceleratorConfig(split_batches=False, dispatch_batches=None, even_batches=True, use_seedable_sampler=True, non_blocking=False, gradient_accumulation_kwargs=None, use_configured_state=False)
• parallelism_config: None
• deepspeed: None
• label_smoothing_factor: 0.0
• optim: OptimizerNames.ADAMW_TORCH_FUSED
• optim_args: None
• adafactor: False
• group_by_length: False
• length_column_name: length
• report_to: ['wandb']
• project: huggingface
• trackio_space_id: trackio
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: None
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• skip_memory_metrics: True
• use_legacy_prediction_loop: False
• push_to_hub: False
• resume_from_checkpoint: None
• hub_model_id: None
• hub_strategy: HubStrategy.EVERY_SAVE
• hub_token: None
• hub_private_repo: None
• hub_always_push: False
• hub_revision: None
• gradient_checkpointing: True
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• include_for_metrics: []
• eval_do_concat_batches: True
• fp16_backend: auto
• push_to_hub_model_id: None
• push_to_hub_organization: None
• push_to_hub_token: None
• _n_gpu: 1
• mp_parameters:
• auto_find_batch_size: False
• full_determinism: False
• torchdynamo: None
• ray_scope: last
• ddp_timeout: 1800
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• include_tokens_per_second: False
• include_num_input_tokens_seen: no
• neftune_noise_alpha: None
• optim_target_modules: None
• batch_eval_metrics: False
• eval_on_start: False
• use_liger_kernel: False
• liger_kernel_config: None
• eval_use_gather_object: False
• average_tokens_across_devices: True
• model_init_kwargs: None
• chat_template_path: None
• dataset_text_field: text
• dataset_kwargs: None
• dataset_num_proc: None
• eos_token: None
• pad_token: None
• max_length: 1024
• packing: False
• packing_strategy: bfd
• padding_free: False
• pad_to_multiple_of: None
• eval_packing: None
• completion_only_loss: None
• assistant_only_loss: False
• loss_type: nll
• activation_offloading: False
• vllm_sampling_params: None
• unsloth_num_chunks: -1
• unsloth_logit_chunk_multiplier: None
• unsloth_grpo_mini_batch: None
• max_seq_length: 29000
• model_name: Snowflake/Arctic-Text2SQL-R1-7B
• train_batch_size: 1
• val_batch_size: 1
• num_epochs: 2
• lora_use_rslora: True
• lora_r: 64
• lora_alpha: 128
• lora_dropout: 0.1

Train Dataset Example

<|im_start|>system
You are an expert in SQL++ query generation. Given a document schema and a natural language query, generate a valid SQL++ query.

Task Instructions:
- Backtick-quote field names that are reserved keywords or contain spaces/special characters.
  WRONG: SELECT value, Enrollment (K-12) ...
  RIGHT: SELECT `value`, `Enrollment (K-12)` ...

- SUBSTR is 0-based: SUBSTR(str, 0, 4) returns the first 4 characters. Use this for year extraction from date strings.
  WRONG: SUBSTR(dob, 1, 4) = '1990'
  RIGHT: SUBSTR(dob, 0, 4) = '1990'

- Only use keyspaces and fields present in the schema; do not infer array, object, or foreign-key structure unless the schema shows it.
  WRONG: UNNEST t.tags AS tag (when `tags` is a plain string in schema)
  RIGHT: WHERE t.tags = 'sports'

- Never use CAST(); it is not supported in SQL++.
  WRONG: CAST(price AS FLOAT)
  RIGHT: TO_NUMBER(price)

- Use the exact field named in the question; do not substitute a related variant.
  WRONG: question asks for `revenue`, query uses `total_sales`
  RIGHT: query uses `revenue`

- When similar fields exist, prefer the one whose name most literally matches the question; use sample values to distinguish (e.g., `type` vs `types`, `id` vs `uuid`).
  WRONG: question asks for "account type", query uses `types` (samples: [1,2,3])
  RIGHT: uses `type` (samples: ["savings","checking"])

- Prefer a direct count or pre-aggregated field over computing it from related records when one exists.
  WRONG: (SELECT COUNT(*) FROM reviews r WHERE r.product_id = p.id) >= 3
  RIGHT: WHERE p.review_count >= 3

- Wrap string fields in TO_NUMBER() before numeric aggregation or ordering.
  WRONG: AVG(p.score)  when score is stored as "8.5"
  RIGHT: AVG(TO_NUMBER(p.score))

- If one collection contains all needed fields and filters, do not join.
  WRONG: FROM orders o JOIN orders o2 ON ...
  RIGHT: FROM orders o WHERE o.status = 'shipped'

- Use DISTINCT when unique values are requested or when a join could produce duplicates.
  WRONG: SELECT c.id FROM customers c JOIN orders o ON c.id = o.customer_id
  RIGHT: SELECT DISTINCT c.id ...

- For yes/no questions, return a single existence answer, not matching rows.
  WRONG: SELECT e.name FROM employees e WHERE e.dept = 'HR'
  RIGHT: SELECT COUNT(*) > 0 FROM employees e WHERE e.dept = 'HR'

- When listing entities with no specified attribute, return the entity identifier.
  WRONG: question says "list employees", query returns SELECT e.name
  RIGHT: SELECT e.id

- No colon after FROM.
  WRONG: FROM: orders o
  RIGHT: FROM orders o

- Every alias in a statement must be unique. Couchbase does not allow the same alias to be assigned more than once, even across subqueries or when referencing the same collection.
  WRONG: SELECT * FROM orders o WHERE o.id IN (SELECT RAW o.ref_id FROM orders o WHERE ...)
  RIGHT: SELECT * FROM orders o WHERE o.id IN (SELECT RAW o2.ref_id FROM orders o2 WHERE ...)

- Match literal types to schema field types; quote string-typed fields even when values look numeric.
  WRONG: WHERE zip_code = 10001  (zip_code type is string in the schema)
  RIGHT: WHERE zip_code = "10001"
<|im_end|>
<|im_start|>user

Bucket Name: bird_training_bucket
Scope Name: restaurant

Collection Schema:
{"`bird_training_bucket`.`restaurant`.`location`": {"Flavor": "", "properties": {"id_restaurant": {"samples": [1534], "type": "number"}, "city": {"samples": [["berkeley"]], "type": "string"}, "street_name": {"samples": [["addison st"]], "type": "string"}, "street_num": {"samples": [[427]], "type": "number"}}, "type": "object"}, "`bird_training_bucket`.`restaurant`.`generalinfo`": {"Flavor": "", "properties": {"city": {"samples": ["emeryville"], "type": "string"}, "review": {"samples": [2], "type": "number"}, "label": {"samples": ["doucet's lounge"], "type": "string"}, "id_restaurant": {"samples": [196], "type": "number"}, "food_type": {"samples": ["american"], "type": "string"}}, "type": "object"}, "`bird_training_bucket`.`restaurant`.`geographic`": {"Flavor": "", "properties": {"city": {"samples": ["charlotte"], "type": "string"}, "region": {"samples": ["bay area"], "type": "string"}, "county": {"samples": ["alameda county"], "type": "string"}}, "type": "object"}}

Natural Language Query:
Show the identifying street address details for places that serve alcohol (a “bar”), located in Oakland, with a review score of 2.7.
<|im_end|>
<|im_start|>assistant

```sql++
SELECT T2.`street_num` FROM `bird_training_bucket`.`restaurant`.`generalinfo` AS T1 INNER JOIN `bird_training_bucket`.`restaurant`.`location` AS T2 ON T1.`id_restaurant` = T2.`id_restaurant` WHERE T1.`review` = 2.7 AND T2.`city` = 'oakland' AND T1.`food_type` = 'bar'

<|im_end|>