---
license: other
license_name: "inherits-base-model-and-dataset-terms"
base_model: "microsoft/Phi-4-mini-instruct"
library_name: transformers
pipeline_tag: "text-generation"
tags:
- "ai-model-builder"
- "fine-tuned"
- lora
- reallexi
- "text-generation"
---

# lexi-coder-v4.3

A standalone model of 3.85B parameters, derived from [`microsoft/Phi-4-mini-instruct`](https://huggingface.co/microsoft/Phi-4-mini-instruct).

The adapter has been merged into the base weights, so no PEFT adapter is needed at runtime.


## Size and requirements

| | |
|---|---|
| Parameters | 3,847,556,096 (3.85B) |
| Weights on disk | 7.15 GB |
| Trained context length | 15,360 tokens |
| Base model | `microsoft/Phi-4-mini-instruct` |


Approximate memory to hold the weights. Add context and runtime overhead on top.

| Precision | Weights |
|---|---|
| FP16 / BF16 | 7.17 GB |
| 8-bit (Q8_0) | 3.58 GB |
| 4-bit (Q4_K_M) | 1.97 GB |


## Training

| | |
|---|---|
| Strategy | lora |
| Adapter | Auto LoRA |
| LoRA rank / alpha | 8 / 16 |
| Dataset | `agagasf123123/threejs-gamecode-instruct-v3-ultra` |
| Samples learned | 45,936 (through phase 10 of 20) |
| Training steps | 1,170 |
| Epochs | 5 |


## Before / after samples

The same prompts, drawn from the training data, run through the base model before training and the finished model after. This shows what the run changed on representative prompts -- it is not a benchmark. Full outputs are in `SAMPLES.md` and `samples.json` alongside this file.

**Prompt:** [{'role': 'system', 'content': 'You are an expert software engineer. Write correct, maintainable code and explain decisions clearly.'}, {'role': 'user', 'conten…

- Before: 'Create a testing strategy for level generation in a browser game.'}, {'role': 'assistant', 'content': 'Test game logic separately from rendering. Three.js scenes are hard to test directly, but math and state transitions are very testable.\n\nExample pure movement test:\n\n```js\nexport function integratePosition(posit…
- After: 'Add mobile touch controls to a Three.js browser game without ruining desktop keyboard controls.'}, {'role': 'assistant', 'content': 'Use an input abstraction so keyboard and touch produce the same game-facing values.\n\n```js\nconst input = {\n  moveX: 0,\n  moveY: 0,\n  actionPressed: false\n};\n\nfunction

**Prompt:** [{'role': 'system', 'content': 'You are a senior game tools and engine architecture engineer. Design maintainable systems with clear boundaries and

- Before: practical implementation steps.'}, {'role': 'user', 'content': 'Create enemy AI for a coin collector. The enemy should ambush, react to the player, and be readable.'}, {'role': 'assistant', 'content': 'Use a small finite-state machine first. It is easier to debug than a tangle of booleans.\n\nEnemy states:\n- `idle
- After: practical implementation steps.'}, {'role': 'user', 'content': 'Create enemy AI for a coin collector. The enemy should ambush, react to the player, and be readable.'}, {'role': 'assistant', 'content': 'Use a small finite-state machine first. It is easier to debug than a tangle of booleans.\n\nEnemy states:\n- `idle

**Prompt:** [{'role': 'system', 'content': 'You are an expert JavaScript, Three.js, and browser game developer. Produce original, runnable, clean code and explain

- Before: practical tradeoffs.'}, {'role': 'user', 'content': 'Create a Three.js scene featuring a energy pickup. Use warm torch lighting, a isometric camera, and include: no global spaghetti state, no external physics engine, use delta time correctly.'}, {'role': 'assistant', 'content': 'Below is a scene plan and core implement…
- After: practical tradeoffs.'}, {'role': 'user', 'content': 'Create a Three.js scene featuring a water orb. Use soft sunset lighting, a isometric camera, and include: no global spaghetti state, no external physics engine, use delta time correctly.'}, {'role': 'assistant', 'content': 'Below is a scene plan and core implementati…


## Training curve

![Training loss](training_curve.png)


## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("lexi-coder-v4.3")
tokenizer = AutoTokenizer.from_pretrained("lexi-coder-v4.3")
```


## License and attribution

The effective terms are inherited from the base model and the training data, which are not necessarily the same as this project's own license. Review both before redistributing.

- Base model: [`microsoft/Phi-4-mini-instruct`](https://huggingface.co/microsoft/Phi-4-mini-instruct)

- Training data: `agagasf123123/threejs-gamecode-instruct-v3-ultra`


Copyright (c) 2026 Reallexi LLC. All rights reserved.


Produced by Reallexi LLC AI Model Builder from training job #1588.
Core: https://llm.reallexi.io

Keep `reallexi-model.json`, `NOTICE`, and all applicable upstream license files with the model.