keisuke-miyako/springsea-0.1

A fine-tuned code generation model for 4D programming

Overview

SpringSea is a fine-tuned version of google/gemma-4-E2B-it, specialized for generating, completing, and reasoning about 4D code — the proprietary language powering 4D application development.

The goal of this project is to develop a coding assistant model that understands the full 4D development ecosystem.

Training Details

Property	Details
Base Model	google/gemma-4-E2B-it
Fine-tune Method	LoRA
Primary Language	4D
Context Window	128K tokens
License	Apache 2.0

---

Type	Rows	Size	LoRA rank	LoRA alpha	Warmup ratio	Learning rate	Epochs	Steps	Duration
Synthetic ChatML	4750	15.51 MB	32	64	0.05	5e-5	2	2:20:49	594

Changes

• Rank: 64→32
• Learning rate: 1e-4→5e-5
• Epoch: 3→2

---

0.1.6 Hallucinates; FAIL💥

Type	Rows	Size	LoRA rank	LoRA alpha	Warmup ratio	Learning rate	Epochs	Steps	Duration
Synthetic ChatML	3731	11.81 MB	32	128	0.05	1e-4	3	702	2:42:32

0.1.5

Type	Rows	Size	LoRA rank	LoRA alpha	Warmup ratio	Learning rate	Epochs	Steps	Duration
Synthetic ChatML	3106	7.49 MB	64	128	0.05	1e-4	3	585	2:17:01

0.1.4

Type	Rows	Size	LoRA rank	LoRA alpha	Learning rate	Epochs	Steps
Synthetic ChatML	1645	4.42 MB	64	128	1e-4	3	309

0.1.3

Type	Rows	Size	LoRA rank	LoRA alpha	Learning rate	Epochs	Steps
Synthetic ChatML	1201	3.03 MB	64	128	1e-4	3	228

0.1.2

Type	Rows	Size	LoRA rank	LoRA alpha	Learning rate	Epochs	Steps
Synthetic ChatML	953	1.53 MB	64	128	1e-4	3	180

0.1.1

Type	Rows	Size	LoRA rank	LoRA alpha	Learning rate	Epochs	Steps
Synthetic ChatML	1786	1.16 MB	64	128	2e-5	2	894

0.1.0

Type	Rows	Size	LoRA rank	LoRA alpha	Learning rate	Epochs	Steps
Synthetic ChatML	1786	1.16 MB	64	128	2e-5	1	447

Instructions

Use this system prompt verbatim:

You are "LEPECQ-0.0.1", an expert of the 4D (4th Dimension) programming language and application development tool.

Followed by a user prompt about 4D.

The models should gravitate towards domain specific thinking when the conversation is about 4D.

This is because the model was fine-tuned using the pattern below:

# system
You are "LEPECQ-0.0.1", an expert of the 4D (4th Dimension) programming language and application development tool.

# user 
I have a [Contacts] table with 80k rows. I need all contacts from Paris sorted by last name. Is there a faster way than looping with a classic query?

# model
<|think|>
This is an ORDA vs classic query performance question, INTERMEDIATE tier. The gotcha is that orderBy() on an entity selection is lazy and deferred, so I should show the full chain and also m

---

Limitations

• Outputs should be reviewed by a 4D developer before use in production

---

License & Attribution

This model is released under the Apache 2.0 License.

Built on google/gemma-4-E2B-it by Google DeepMind, also Apache 2.0.

---