Comulative

KAT-Coder-V2.5_JKL-Luau-NVFP4

README

License: apache-2.0

Intended uses

Primary: local / self-hosted Roblox Luau coding assistant and task executor:

Idiomatic Luau modules (--!strict, services, remotes, DataStore patterns)
Roblox client/server split, RemoteEvents / RemoteFunctions, replication-aware snippets
Code completion / continuation on existing Luau scripts
Roblox engine Q&A grounded in specialty SFT data

Recommended deployment pattern: use a stronger planning / review model for architecture and security, and this model as a fast local executor for Luau implementation (low latency, private weights, Blackwell-friendly NVFP4).

Not intended for: unsupervised production game economy / anti-cheat design without human review; non-Roblox general agenting as a drop-in frontier replacement; vision/multimodal tasks (language-only lineage).

What we did (training pipeline)

1. Base

Started from KAT-Coder-V2.5-Dev (Qwen3.5 MoE coding model, language-only open weights; vision tower declared in config but not shipped upstream).

2. Domain SFT data mix

Built a chat-formatted SFT mix (42,302 train rows after filtering/dedup; ~57M tokens) from public Hugging Face datasets (local mirror under data/).

Hugging Face datasets used in training

Table with columns: Hub dataset, Role in mix, Approx. SFT rows (source tag)
Hub dataset	Role in mix	Approx. SFT rows (source tag)
TorpedoSoftware/Roblox-Luau-Reasoning-v1.0	Prompt → CoT + Luau code + explanation (`train`)	14,840 `luau-reasoning`
Pinkstack/luaucoder-instructions-v3-SFT	Instruction SFT (filtered; cap ~8k quality rows)	7,927 `pinkstack-sft`
khtsly/luau-stack-hq

Total train examples: 42,302 (plus 400 held-out mix rows for training-time val).

Evaluation dataset (not used for gradient updates)

Table with columns: Hub dataset, Use
Hub dataset	Use
TorpedoSoftware/RobloxQA-v2.0 `test` (3,000 MCQ)	Held out for baseline / future eval only

Downloaded but not used in this SFT mix

Table with columns: Hub dataset, Notes
Hub dataset	Notes
TorpedoSoftware/roblox-info-dump	Roblox/Luau docs scrape present under `data/`; not mixed into the epoch-1 SFT JSONL

Formatting used the base model chat template, with prompt tokens masked (train on completions). Optional system prompts mixed in (~30% Roblox-assistant style). Decompiled-looking completion snippets filtered out. Pinkstack rows required code fences + minimum reasoning length.

3. LoRA fine-tune (1 epoch)

Table with columns: Hyperparameter, Value
Hyperparameter	Value
Method	LoRA (PEFT), bf16 base
Rank / alpha	r=64, α=128, dropout 0.05
Targets	Attention (q/k/v/o), linear-attn projections, shared-expert MLP; not per-routed experts
Context	4096
Effective batch	16 (microbatch 2 × grad accum 8)
Epochs shipped here	1 (stopped at step 2642 / 5284 of a 2-epoch schedule)
Optim	AdamW fused, LR 1e-4 cosine, warmup 40
Hardware	1× RTX PRO 6000 96GB, Windows 11 + WSL2 Ubuntu

A full Trainer checkpoint (adapter + optimizer) was frozen at epoch 1 so a second epoch can be resumed later if desired. This Hub repo is the merged + NVFP4 product of that epoch-1 adapter, not the raw LoRA.

4. Merge & cleanup

Merged LoRA into base bf16 weights
Stripped randomly initialized visual.* keys so the checkpoint matches the language-only upstream release

5. NVFP4 quantization

Tooling: llm-compressor oneshot PTQ
Scheme: NVFP4 (W4A4, nvfp4-pack-quantized)
Calibration: 256 samples @ 4096 from the Luau SFT mix
moe_calibrate_all_experts=True
Ignores: lm_head, visual towers, router gates, embeddings, linear-attn (see recipe.yaml)

Why NVFP4: native-friendly 4-bit float path for Blackwell inference stacks (e.g. vLLM on modern NVIDIA data-center / pro GPUs). Smaller footprint (~22GB weights here) vs full bf16 (~70GB class).

Base model evaluation note

Unmodified base KAT-Coder-V2.5-Dev on RobloxQA-v2.0 test (3000 questions), MMLU-style log-prob forced choice, bf16, HF Transformers:

Baseline: 87.60% (2628 / 3000) — measured 2026-08-01.

Post–fine-tune / post-NVFP4 RobloxQA numbers for this checkpoint may be published later; treat the above as the starting point the specialty run was designed to preserve or beat on knowledge while improving Luau generation.

How to use

Transformers / llm-compressor compressed load

Follow current llm-compressor / Transformers docs for NVFP4 compressed-tensors checkpoints. Ensure a stack that understands quantization_config with format nvfp4-pack-quantized.

vLLM (Blackwell)

Prefer a recent vLLM build with NVFP4 + Qwen3.5 MoE support. Language-only serving flags may still apply (upstream was language-only; config may still mention vision). Example pattern (versions change quickly—check your vLLM release notes):

bash
vllm serve Comulative/KAT-Coder-V2.5_JKL-Luau-NVFP4 \
  --quantization modelopt_fp4 \
  # plus whatever your vLLM build requires for NVFP4 / Qwen3.5 MoE

Use the model’s chat template (chat_template.jinja / tokenizer config). For agentic coding UIs, Qwen-style reasoning / tool parsers may apply depending on runtime.

Prompting tips

Prefer narrow tasks: one module, one bugfix, one remote contract
State client vs server, remote names, and data shapes
Ask for --!strict Luau when you want typed modules
Review security (RemoteEvent trust, DataStore races) with a stronger planner model before shipping

Training summary card

text
Base:     Kwaipilot/KAT-Coder-V2.5-Dev (Qwen3.5-MoE ~35B-A3B)
SFT:      LoRA r=64, 1 epoch, Luau/Roblox mix ~42k examples
Merge:    bf16 merge + strip visual.*
Quant:    NVFP4 (llm-compressor), 256 calib samples, all MoE experts
Trainer:  @dylanjkl / Comulative Limited (UK)
GPU:      1× NVIDIA RTX PRO 6000 Blackwell 96GB

Limitations & risks

Domain SFT can narrow style; may be weaker than base on non-Roblox agentic coding
1 epoch only in this release; second epoch optional, not included
NVFP4 is a lossy quant vs bf16 merge—validate on your tasks
No Studio execution loop in training—always test Luau in Roblox
May produce insecure remote/DataStore patterns; do not deploy unreviewed
MoE + hybrid linear attention runtimes are still evolving—pin tested engine versions

License

Weights are a derivative of Kwaipilot/KAT-Coder-V2.5-Dev. Unless otherwise required by the base model license, this distribution is provided under Apache-2.0. Review the base model card and license for any additional terms.

Citation / attribution

text
Base model: Kwaipilot/KAT-Coder-V2.5-Dev
Fine-tune & NVFP4 release: Comulative Limited (UK) / @dylanjkl
Hub: Comulative/KAT-Coder-V2.5_JKL-Luau-NVFP4
Hardware: 1× NVIDIA RTX PRO 6000 Blackwell 96GB

Maintainers

@dylanjkl — Comulative Limited (United Kingdom)
Organization: Comulative

For issues with this fine-tune/quant, contact the Comulative maintainers—not Kwaipilot—unless the bug is reproducible on the unmodified base.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider