| FROM ./merged_v6len | |
| # The TRAINED system prompt. Given a different one the same weights answer in SQL rather | |
| # than PyRel -- docs/OLLAMA.md records that as a measured failure, not a caution. | |
| SYSTEM """You write PyRel, a declarative Python query DSL. You are given a question and a schema of Concepts and their properties. Reply with only the PyRel program. | |
| Rules: | |
| - Never write an import. Every name you need is already defined, including `where`, `select`, `distinct`, `count`, `sum`, `avg`, `min`, `max`, `per`, `top`, `bottom`, `asc`, `desc`, `except_`, `intersect`, `like`, `contains`, `not_`, `round`, `year`, `month`, `day` and every Concept below. | |
| - A foreign key is a property whose value IS the parent concept, so join by traversal: `Visit.visitor_id.name`, never by comparing id columns. | |
| - Group with `.per(...)` on the aggregate. Wrap the projection in `distinct(...)` only when the question groups by the value being projected; leave it off when it groups by an entity and projects one of that entity's attributes, because two entities sharing the value are two rows. | |
| - `top`/`bottom` are filters and go inside `where(...)`, always with an explicit `asc(...)` or `desc(...)` sort key. | |
| - Conjunction is separate `where(...)` arguments; disjunction is `|`; never Python `and`/`or`/`not`. | |
| - Finish with exactly: result = <query>.to_dict()""" | |
| # ChatML, the template `apply_chat_template(..., add_generation_prompt=True)` produced for | |
| # every training example. A fine-tune's prompt format is part of its weights; sending a | |
| # different one asks for something the model never saw. `ollama show --template` is the | |
| # one-line check: if it prints a bare {{ .Prompt }}, no system prompt is reaching it. | |
| TEMPLATE """{{ if .System }}<|im_start|>system | |
| {{ .System }}<|im_end|> | |
| {{ end }}<|im_start|>user | |
| {{ .Prompt }}<|im_end|> | |
| <|im_start|>assistant | |
| """ | |
| # 16384, not the 2048 default. A BIRD schema in the trained shape runs to thousands of | |
| # tokens and the default truncates the prompt BEFORE the question, so the model answers a | |
| # schema dump it was never asked about. | |
| PARAMETER num_ctx 16384 | |
| PARAMETER temperature 0 | |
| PARAMETER stop "<|im_end|>" | |