caveman-1

caveman-1 is a LoRA adapter for Qwen3.8-27B that makes the model answer in terse caveman speech with a pirate accent. It is packaged as a .ninfer adapter for NInfer-4090 and loads per request beside the base weights. It is not a PEFT, Transformers, or GGUF file.

Hrrm. Register renaming give each write its own private register, so later read grab right value
and no stall wait on old one. It clear most false hazards, matey, and let out-of-order engine run
ahead of program order.

File

Field	Value
Filename	caveman-1.lora.ninfer
Size	84,471,808 bytes
SHA-256	1c10e3d042f7a37e32f4867d2e76e8baea6898c4452e0f3e32f05396136971c9
Kind	generative adapter
Rank	16, alpha 32, folded into B
Sites	query, output gate, key, value, and attention output on the 16 attention layers, Gated DeltaNet output on the other 48, MLP down projection on all 64
Parameters	42,205,184

Use it

The adapter needs NInfer-4090 and the Qwen3.8-27B NInfer artifact. Download the adapter into the directory that the server reads with --lora-dir:

hf download tensorninja2/caveman-1 caveman-1.lora.ninfer --local-dir lora
./build-sm89/apps/ninfer-serve models/qwen3_8_27b.ninfer --lora-dir lora --lora-slots 2

The server exposes the adapter as the model qwen3.8-27b-caveman-1 on OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages:

curl -s http://127.0.0.1:8080/v1/chat/completions -H 'content-type: application/json' -d '{
  "model": "qwen3.8-27b-caveman-1", "enable_thinking": false, "temperature": 0,
  "messages": [{"role": "user", "content": "Explain register renaming in two sentences."}]}'

Requests for the base model and for other adapters can run in the same decode batch. See Load LoRA adapters per request.

Best with Caudra

Any agent that speaks the OpenAI or Anthropic APIs can use the adapter by its model name. For the most native experience, use Caudra, the coding agent from the author of NInfer-4090. With the ninfer-4090 provider from Use it from Caudra, Caudra discovers the adapter, and you select it as ninfer-4090/qwen3.8-27b-caveman-1 in /model.

Training

QLoRA on an NF4-quantized Qwen3.8-27B base: PEFT modules q_proj, k_proj, v_proj, o_proj, down_proj, and out_proj, 120 steps, maximum sequence length 1,024, completion-only loss, and thinking off. The corpus is the caveman_pirate SFT split of the llm-datasets repository. tools/convert/qwen3_8_27b/convert_lora.py converted the PEFT directory to .ninfer.

Limits

• The adapter changes the register of the answer. It was not trained or evaluated for task accuracy.
• NInfer's MTP draft head is not adapted, so speculative decoding accepts fewer draft tokens with this adapter than with the base weights.
• It was trained with thinking off.

License

Apache-2.0, as is the Qwen3.8-27B base.