🧠 imagine-v10

100% on held-out. Zero hallucinations. Analytical SQL that actually works.

Natural language in. Correct PostgreSQL out. ~1B parameters. No GPU. No API key. No metered inference.

Part of project imagine — Interchained

---

The failure that built v10

v9 was asked for a LEFT JOIN analytical query:

-- "List every customer, their total completed spending, sort by spending desc"

It returned:

SELECT c.id, c.name, COALESCE(o.total, 0) AS total_spending
FROM customers c
LEFT JOIN orders o ON c.id
-- ...truncated. Dead.

The correct answer:

SELECT c.id, c.name, COALESCE(SUM(o.total), 0) AS total_spending
FROM customers c
LEFT JOIN orders o ON c.id = o.customer_id AND o.status = 'completed'
GROUP BY c.id, c.name
ORDER BY total_spending DESC, c.id ASC;

v9 couldn't do multi-clause analytical queries. v10 can.

---

🎯 What v10 adds

Where v9 added write capability (INSERT/UPDATE/DELETE), v10 adds complex analytical queries:

• LEFT JOIN + GROUP BY + COALESCE — the exact v9 failure mode, now handled
• HAVING filters — post-aggregation conditions
• Subquery comparisons — nested analytical logic
• Complex ORDER BY — multi-column, expression-based sorting
• Multi-table JOINs — beyond simple two-table patterns

---

📊 The numbers

Model	Held-out (telemetry)	Protocol	Execute rate	Invented cols
imagine-v8	50.0% (20/40)	100%	—	—
imagine-v9	90.0% (36/40)	100%	97.5%	0
imagine-v10	100.0% (40/40)	100%	100%	0

Perfect score. Zero wrong answers. Zero hallucinations. Zero refusals.

---

🧬 How v10 was built

Two-stage fine-tuning from Interchained/imagine-v9:

Stage 1 — Full fine-tune (smoke)

• Corpus: 3,216 admitted pairs (99.94% gate admit rate)
• New: Analytical query templates (LEFT JOIN aggregations, HAVING, subqueries)
• Epochs: 3, LR 1e-5, cosine schedule
• Final loss: 0.0052 (v9 was 0.0320 — 6x better)
• Time: 4.1 min on H200, 26k tok/s

Stage 2 — LoRA refinement

• Base: v10-smoke checkpoint
• Rank: 16, LR 5e-6
• Epochs: 3
• Final loss: 0.0244
• Time: 5.4 min on H200
• Adapter folded into full checkpoint

The corpus

3,218 candidates forged, 3,216 admitted:

Schema	Templates	Writes	Total
shop	356	388	744
clinic	545	290	835
library	481	257	738
fleet	297	226	523
audit	252	126	378

Nothing enters training that a live database hasn't agreed with.

---

🔬 The execution gate

Every candidate goes through L0–L4:

candidate SQL ──► L0  real PostgreSQL parser     not a regex
                 L1  read-only + bounded          no writes, no sleeps
                 L2  EXPLAIN on live schema       hallucinations die HERE
                 L3  execute, timed, capped       real rows
                 L4  SAME ANSWER as ref?          ◄── the one that matters
                                              ▼
                                       admitted to corpus

68/68 gate tests passing. The gate is the truth authority — not a bigger model, not vibes.

---

📦 Output protocol

SQL wrapped in sentinel blocks:

<<<SQL>>>
SELECT c.id, c.name, COALESCE(SUM(o.total), 0) AS total_spending
FROM customers c
LEFT JOIN orders o ON c.id = o.customer_id AND o.status = 'completed'
GROUP BY c.id, c.name
ORDER BY total_spending DESC, c.id ASC;
<<<END>>>

Can also refuse:

block	meaning
<<<SQL>>>	here is your query
<<<UNANSWERABLE>>>	this schema cannot answer that
<<<CLARIFY>>>	ambiguous — here's what's missing

⚠️ A truncated generation is not an answer. An unterminated block extracts to nothing.

---

💻 Loading v10

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Interchained/imagine-v10"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# Ask for an analytical query
prompt = """Schema:
customers(id, name)
orders(id, customer_id, total, status)

List every customer ID, name, and total completed spending.
Include customers with no completed orders (show 0).
Sort by spending descending, then ID ascending."""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(inputs["input_ids"], max_new_tokens=256, do_sample=False)

print(tokenizer.decode(output[0], skip_special_tokens=True))

---

🎯 Identity

The deployed model identity is Imagine.

Built and fine-tuned by Interchained.

DeepSeek-Coder is part of the upstream lineage (via v8 → v9 → v10), but the deployed identity is Imagine.

---

⚡ Local-first

Runs on your hardware. No API key. No metered inference. No cloud dependency.

---

📐 The rules

1 · Don't write a verifier — the engine already shipped one. Real parser. Real planner. Real rows.

2 · Assert the property, not a proxy. Execution accuracy, not string similarity.

3 · Schema goes in the prompt, not in the weights. The model learns "read the schema you were handed" — not memorise ours.

---

⚠️ Limitations

v10 is a research checkpoint. It may still:

• generate incorrect SQL on novel patterns
• misunderstand ambiguous requests
• produce writes with wrong WHERE clauses — always review before executing

Generated SQL should be reviewed before use in production. This applies doubly to writes.

---

🔒 Security

Do not rely on model behavior alone for database safety. Production systems should enforce:

• least-privilege database roles
• statement timeouts and row limits
• query validation and schema restrictions
• application-level authorization
• audit logging
• human approval for all write statements

---

Built by Interchained · ownership at every layer, including the model

3 > 1 — Mark drives, oracle points, Muse builds