Spaces:
Paused
Paused
| [client] | |
| name = "sqlbot" | |
| task = "sql" | |
| description = "A BI startup wants a small text to SQL model that runs on a CPU box next to their warehouse, answers with one SQL statement and nothing else, and is judged by whether the query returns the right rows." | |
| [data] | |
| min_chars = 40 | |
| max_chars = 6000 | |
| val_fraction = 0.0 | |
| max_records = 0 | |
| seed = 1 | |
| dedup = "exact" | |
| num_perm = 64 | |
| bands = 8 | |
| shingle = 5 | |
| workers = 4 | |
| [[data.sources]] | |
| name = "gretel_train" | |
| url = "https://huggingface.co/datasets/gretelai/synthetic_text_to_sql/resolve/main/synthetic_text_to_sql_train.snappy.parquet" | |
| format = "parquet" | |
| split = "train" | |
| [[data.sources]] | |
| name = "gretel_test" | |
| url = "https://huggingface.co/datasets/gretelai/synthetic_text_to_sql/resolve/main/synthetic_text_to_sql_test.snappy.parquet" | |
| format = "parquet" | |
| split = "val" | |
| [tokenizer] | |
| vocab_size = 8192 | |
| sample_bytes = 100000000 | |
| [model] | |
| dim = 384 | |
| n_layers = 6 | |
| n_heads = 6 | |
| n_kv_heads = 2 | |
| seq_len = 768 | |
| rope_theta = 10000.0 | |
| [pretrain] | |
| tokens = 80000000 | |
| batch_tokens = 98304 | |
| micro_batch = 32 | |
| lr = 0.001 | |
| min_lr = 0.0001 | |
| warmup_steps = 200 | |
| weight_decay = 0.1 | |
| grad_clip = 1.0 | |
| eval_every = 200 | |
| eval_batches = 20 | |
| checkpoint_every = 200 | |
| log_every = 20 | |
| dtype = "auto" | |
| compile = false | |
| seed = 1 | |
| [sft] | |
| epochs = 2 | |
| lr = 0.0001 | |
| batch_size = 32 | |
| max_examples = 0 | |
| warmup_steps = 100 | |
| [dpo] | |
| beta = 0.1 | |
| lr = 0.00001 | |
| pairs = 4000 | |
| epochs = 1 | |
| batch_size = 8 | |
| candidates = 4 | |
| temperature = 0.8 | |
| warmup_steps = 20 | |
| [generation] | |
| max_new_tokens = 160 | |
| temperature = 0.0 | |
| top_p = 1.0 | |
| top_k = 0 | |
| repetition_penalty = 1.0 | |
| [eval] | |
| samples = 1000 | |
| batch_size = 64 | |
| perplexity_batches = 50 | |
| [[eval.gates]] | |
| metric = "execution_accuracy" | |
| min = 0.4 | |
| [[eval.gates]] | |
| metric = "valid_rate" | |
| min = 0.8 | |
| [task] | |
| timeout_ms = 2000 | |
| max_rows = 1000 | |
| [export] | |
| quantize = "int8" | |
| [serve] | |
| system_prompt = "You translate questions about a database into one SQL statement. Reply with the SQL only." | |
| stop = [";"] | |
| example = "Schema:\nCREATE TABLE orders (id INT, customer TEXT, amount DECIMAL(10,2), placed_on DATE);\nINSERT INTO orders VALUES (1, 'Ana', 120.50, '2024-01-05'), (2, 'Ben', 80.00, '2024-02-11'), (3, 'Ana', 45.25, '2024-02-20');\n\nQuestion: What is the total order amount per customer?" | |
| max_batch = 8 | |
| batch_wait_ms = 20 | |