This repository contains the GGUF inference variants of PB Qwen3.5-9B DoRa v7 Stable, a fine-tuned specialist model for PowerBASIC 11 (PBWIN) code generation, syntax assistance, and inline-assembly help.

Model description

Base model: Qwen/Qwen3.5-9B
Fine-tuning method: DoRa (Weight-Decomposed LoRA) + 4-bit quantization ("QDora") via Unsloth
Domain: PowerBASIC 11 language, Win32 API bindings, and x86 inline assembly
Model family: PB v7 Stable (May 2026)

Included files

File	Quantization	Size	Use case
`pb_qwen3_5_9b_dora_v7_stable_bf16.gguf`	BF16	16.7 GB	Maximum quality, GPU only
`pb_qwen3_5_9b_dora_v7_stable_f16.gguf`	F16	16.7 GB	High quality, GPU only
`pb_qwen3_5_9b_dora_v7_stable_q8_0.gguf`	Q8_0	8.9 GB	Near-lossless, fast inference
`pb_qwen3_5_9b_dora_v7_stable_q6_k.gguf`	Q6_K	6.9 GB	Excellent quality / size ratio
`pb_qwen3_5_9b_dora_v7_stable_q4_k_m.gguf`	Q4_K_M	5.2 GB	Recommended default
`pb_qwen3_5_9b_dora_v7_stable_q4_k_s.gguf`	Q4_K_S	5.0 GB	Small, good for CPU offloading
`pb_qwen3_5_9b_dora_v7_stable_q4_0.gguf`	Q4_0	5.0 GB	Smallest, fastest on CPU

Recommendation: Use Q4_K_M for daily GPU inference (RTX 3090/4090/5090). Use Q8_0 or BF16 when absolute fidelity matters.

Training setup

Parameter	Value
DoRa rank (`r`)	96
DoRa alpha (`alpha`)	192
Dropout	0.05
RSLora	Yes
Sequence length	3072
Batch size	1 × 8 gradient accumulation (effective 8)
Max steps	3,000
Learning rate	6.0e-5
LR scheduler	Cosine with warmup
Warmup steps	240
Optimizer	AdamW 8-bit
Dtype	BF16

Dataset

Train: pb_v7_final_PERFECT_plus_asm_train.jsonl — 14,801 examples
Validation: pb_v7_final_PERFECT_plus_asm_val.jsonl — 1,570 examples
Format: ShareGPT conversation (messages array with system / user / assistant turns)
Coverage: PB core commands, string / file / graphics / math functions, Win32 API declarations, and x86 inline-assembly snippets

Training outcomes

Metric	Value
Final step	3,000
Final loss (last chunk avg)	~0.0048
Trainable parameters	175,620,096 (1.83 % of 9.58 B)
Adapter size	~670 MB
Epochs	~1.62
Total runtime	~45 hours (chunked, RTX 5090)

Hardware and runtime

GPU: NVIDIA RTX 5090 32 GB VRAM
OS: Windows 11 (native, WDDM)
CUDA: 12.8 (Blackwell / sm_120)
Framework: PyTorch 2.7+cu128, Unsloth, PEFT, TRL

Training was performed in 50-step chunks to defeat the Windows WDDM memory-leak degradation that affects long-running CUDA processes on consumer GeForce cards. Each chunk restarted the Python process, flushed VRAM completely, and resumed from the latest checkpoint. This kept step times stable at ~25–30 s/step throughout the entire run.

Intended use

PowerBASIC 11 code generation (SUBs, FUNCTIONs, TYPEs, macros, inline ASM)
Syntax checking and debugging assistance
Win32 API declaration lookup and generation
x86 inline-assembly explanation and generation
Conversion of C/C++ snippets to PowerBASIC

Inference

llama.cpp / LM Studio

Download your preferred quantization (e.g., pb_qwen3_5_9b_dora_v7_stable_q4_k_m.gguf).
In LM Studio, load the model and set:
- Context length: 3072 (or up to 8192 if your VRAM allows)
- Stop sequences: ###STOP### (recommended), <|im_end|>, <|endoftext|>
- GPU offload: Max layers for your VRAM

Example prompt format (chat)

<|im_start|>system
You are a PowerBASIC 11 coding expert. Write correct, compilable PBWIN code.<|im_end|>
<|im_start|>user
Write a SUB that reads a text file line by line and prints each line.<|im_end|>
<|im_start|>assistant

Bias, risks, limitations

This is a narrow-domain specialist. General-knowledge answers depend on the base model; the fine-tune primarily improves PB syntax and idioms.
It is designed for code quality and syntax consistency, not safety-policy alignment.
Do not use for safety-critical decisions without human validation.
Inline-assembly generation is best-effort; always verify with an assembler.

Related models

Model	Repo	Purpose
SPR Qwen3.5-9B DoRa VRAM-safe	`Theogott/spr-qwen3.5-9b-dora-vramsafe-gguf`	SPR (MiniRobot) command family
SPR Adapter	`Theogott/spr-qwen3.5-9b-dora-vramsafe-adapter`	PEFT adapter for SPR

Source

Training repo: QwenPB_FineTune (local)
Adapter checkpoints: models/checkpoints_pb_qwen3_5_9b_dora_v7_stable/
Final adapter: models/checkpoints_pb_qwen3_5_9b_dora_v7_stable/final_adapter/
Conversion: llama.cpp/convert_hf_to_gguf.py + llama-quantize

Generated on 2026-05-07. Model trained by Theogott.

Theogott/pb-qwen3_5-9b-powerbasic-ggufs

PB Qwen3.5-9B DoRa v7 Stable — PowerBASIC GGUF Pack