---
library_name: peft
base_model: Qwen/Qwen3-4B-Instruct-2507
tags:
  - code
  - lora
  - rocm
  - habbo
  - game-server-emulation
  - flash
  - shockwave
  - continued-pretraining
language:
  - en
license: apache-2.0
inference: false
---

# FuseLLM-Instruct-4B-v1

A domain-specialist code model for the **Habbo Hotel ecosystem** — server emulators (Java, C#, PHP, Rust) and Flash/Shockwave client tooling (ActionScript, LiveScript, SWF/DIR reverse engineering). Built by continued pretraining of `Qwen3-4B-Instruct-2507` on a curated corpus of ~85K source files (~211M tokens) drawn from Habbo emulator projects, decompiled client code, and CMS/database dumps.

---

## tl;dr

- **What it is:** a 4B-parameter instruction-tuned code model, LoRA-continued-pretrained on Habbo emulator + retro-client source code.
- **Trained on:** a single consumer GPU — AMD Radeon RX 7900 XTX, 24 GB VRAM, ROCm 7.1.1. bf16 LoRA, no quantization.
- **Runs on:** essentially *any* recent consumer GPU. At 4B params, a Q4_K_M GGUF is ~4.5 GB — it fits comfortably in 8 GB VRAM and will run on 6 GB with a smaller context. The full bf16 merged weights (~14.7 GB) fit in 16 GB. If your card was made in the last ~6 years and has ≥6 GB, it can run FuseLLM-Instruct-4B-v1.

---

## Inference — yes, your GPU can run this

The point of shipping a 4B model is that you don't need a datacenter to use it. The training rig was a 24 GB consumer AMD card; **inference needs far less.**

| Path | Size | Min. VRAM (comfortable) | Notes |
|---|---|---|---|
| **GGUF Q4_K_M (recommended)** | ~4.5 GB | 8 GB (runs on 6 GB, short ctx) | Via Ollama / llama.cpp / LM Studio / KoboldCpp. CPU-only also works — slow but functional. |
| GGUF Q8_0 | ~4.9 GB | 8 GB | Near-lossless; only marginally bigger than Q4_K_M. |
| bf16 merged (full precision) | ~14.7 GB + KV cache | 16 GB | The "no quantization" path. 24 GB leaves room for a long context. |

Concrete examples of cards that run it fine:

- **8 GB:** RTX 3060 / 4060 / 5060, RX 6600 / 7600, Arc A580 — Q4_K_M with 4–8K context.
- **12 GB:** RTX 3060 12GB / 4070 / 5070, RX 6700 XT / 7700 XT — Q4_K_M with long context, or Q8_0 comfortably.
- **16 GB:** RTX 4060 Ti 16GB / 4080 / 5070 Ti, RX 7800 XT / 9070 — bf16 merged fits.
- **24 GB:** RTX 3090 / 4090 / 5090, RX 7900 XTX/XT — bf16 merged with a large context, or Q4 with room to spare.

Linux, Windows, macOS (Metal) all supported through llama.cpp / Ollama. **AMD, NVIDIA, and Intel are all first-class** — the GGUF backend is vendor-agnostic.

### Quick start (Ollama)

```bash
ollama run h4bbo/fusellm-instruct-4b-v1          # once uploaded
# or load the local GGUF:
ollama create fusellm-instruct-4b - Modelfile         # FROM ./fusellm-instruct-4b-v1-Q4_K_M.gguf
ollama run fusellm-instruct-4b
```

### Quick start (transformers, bf16)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("h4bbo/FuseLLM-Instruct-4B-v1")
model = AutoModelForCausalLM.from_pretrained(
    "h4bbo/FuseLLM-Instruct-4B-v1",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
```

---

## Training details

### Hardware

| | |
|---|---|
| GPU | AMD Radeon RX 7900 XTX, 24 GB VRAM (Navi 31) |
| Stack | ROCm 7.1.1, PyTorch 2.13.0+rocm7.1 (`gfx1100` native) |
| Why not CUDA | The rig is AMD-only — the whole recipe below is the ROCm path. No H100, no A100, no datacenter. |

This is a **one-consumer-GPU** training run. The 4B base + bf16 LoRA + gradient checkpointing fits in 24 GB without any quantization (QLoRA was deliberately avoided — bitsandbytes-on-ROCm is still "preview" quality). If you have a single 24 GB card of either vendor, you can reproduce this run.

### Base model

`Qwen/Qwen3-4B-Instruct-2507` — bf16 safetensors, Apache-2.0, ~8 GB. Downloaded from HF Hub (requires `HF_TOKEN`).

### Method — bf16 LoRA continued pretraining

Raw-code continued pretraining (full-sequence causal-LM loss, no instruction pairs).

- **LoRA:** r=32, alpha=64, dropout=0.05, all 7 Qwen3 projections (`q/k/v/o/gate/up/down_proj`), `task_type=CAUSAL_LM`.
- **Optimizer/schedule:** `adamw_torch`, lr=1e-4, cosine, warmup_ratio=0.03, max_grad_norm=1.0, 1 epoch.
- **Batching:** per-device batch 2 × grad-accum 16 = effective 32, `max_length=2048` with `packing_strategy="wrapped"` (~65K tokens/step), ~3,050 steps, ~3–5 h.
- **Precision/memory:** `bf16=True`, `gradient_checkpointing=True` (`use_reentrant=False`), `attn_implementation="sdpa"` (FlashAttention-2 is broken on ROCm), `optim="adamw_torch"` (no bitsandbytes). `PYTORCH_HIP_ALLOC_CONF=expandable_segments:True,garbage_collection_threshold:0.8`, `HSA_OVERRIDE_GFX_VERSION=11.0.0`.
- **Loss:** `completion_only_loss` unset — full-sequence causal-LM on packed raw code (correct for continued pretraining, since there are no instruction/response pairs to mask).
- **Merge:** `merge_and_unload(safe_merge=True)` into the bf16 base → standalone safetensors. `save_peft_format=False` (critical — otherwise the adapter re-attaches on reload).

### Training data

| | |
|---|---|
| Files | 84,925 unique (sha256-deduped; 206,707 duplicates removed; 2.07M minified files skipped) |
| Size | 803.2 MB |
| Tokens | ~211M (est. chars/4) |
| Format | `{"text": <redacted file content>}` — TRL `dataset_text_field="text"` |

By language (files): Java 26,335 · C# 19,048 · PHP 10,969 · LiveScript 7,457 · ActionScript 3,801 · Python 3,023 · JavaScript 2,646 · XML 2,569 · HTML 2,023 · CSS 1,470 · Rust 1,442 · C 1,103 · TypeScript 644 · C++ 517 · SQL 494 · VB 362 · JSON 351 · + Markdown/Gradle/YAML/Scala/Lua.

Sources: 123 Quackster Habbo emulator/tooling repos (incl. private: HorusClient, Kurkku, Aleeda, Icarus variants, cappo-emu, …), `ntuative/RELEASE63…`, `deklol/Shockless`, plus deeply-nested Beta-archive extractions (Debbo, BloodLine, Chocohotel, uberEmu, etc.) — 497 archives / 30 GB unpacked. `.sql` DB dumps (148 MB) are included for now and may be dropped in a later revision.

---

## Intended use & limitations

**Intended:** code completion / Q&A for Habbo server-emulator and retro-client development — packet handling, room/item state, CMS schemas, SWF/DIR reverse engineering, Shockwave Lingo, ActionScript 3 client internals.

**Not intended:** general-purpose chat, math, or non-Habbo code generation. This is a *continued-pretraining* of an instruct model on domain code; it is not a general assistant and was not aligned for safety/RLHF beyond what the base instruct model already had.

**Limitations:**
- 4B parameters — strong on domain pattern-completion, weaker than larger models on multi-file reasoning.

---

## License & data provenance

- **Base model:** `Qwen3-4B-Instruct-2507` — **Apache-2.0**. Fine-tuning and redistribution permitted with attribution. ✅
- **This fine-tune (weights):** released under **Apache-2.0** *conditional on the data licensing below*. The LoRA adapter is small and derivative; the merged model inherits both base and data obligations.
- **Training data:** mixed provenance —
  - Author's own repos (fine).
  - **GPL / various third-party emulators** (PHPRetro/Yifan Lu, uberEmu/Meth0d, Holograph, Icarus, etc.) — GPL-derivative debate applies; a model trained on GPL source is arguably a derivative work.

## Citation

If this model is useful, cite the base and this fine-tune:

```bibtex
@misc{fusellm-instruct-4b-v1,
  title  = {FuseLLM-Instruct-4B-v1: a Habbo ecosystem code model},
  note   = {bf16 LoRA continued pretraining of Qwen3-4B-Instruct-2507},
  year   = {2026},
}
```