caveowl-35b-a3b

caveowl is a bf16 LoRA fine-tune of ornith-ai/Ornith-1.5-35B-A3B (a Qwen3.6-35B-A3B derivative, MIT) that answers in a terse "caveman" register: articles, copulas and filler dropped, facts kept. The style is in the weights, not in a system prompt. The training mix also holds the model's tool calls, code and agent-skill use where stock had them.

This is run 5 of the recipe in sjvrensburg/caveowl, which has the full method, data builders, gate scripts and the record of all five runs. It is a 35B port of Julius Brussee's cavegemma. The author uses it as the everyday assistant model on one AMD Strix Halo machine, behind the pi coding agent.

Same prompt, same server settings, temperature 0:

Prompt: Summarize this in two sentences: The 1986 Chernobyl disaster resulted from a flawed reactor design combined with inadequately trained personnel. [...]

Stock Ornith (44 words): The 1986 Chernobyl disaster was caused by a flawed reactor design and inadequately trained personnel who disabled safety systems during a test. A sudden power surge then triggered explosions that destroyed the reactor core and released massive amounts of radioactive material into the atmosphere.

caveowl (30 words): Chernobyl disaster (1986) caused by flawed reactor design and inadequately trained personnel. Operators disabled safety systems during test, power surge caused explosions, destroyed reactor core, released radioactive material into atmosphere.

Files

File	Size	sha256
Ornith-1.5-35B-caveowl-run5-Q4_K_M.gguf	21,713,462,592 B (21.7 GB)	18948c1313f5192a13fe1428e368e98a18588dcbaa5d667673335829acd3e68d
mmproj-Ornith-1.5-35B-A3B-f16.gguf	899,283,296 B (0.9 GB)	815c9a671991aceeef46ca1aa103b64499db8cfadf21c83005a20ed839d53c7c
adapter/adapter_model.safetensors	3,767,769,408 B (3.77 GB)	316eb48ffbac4474cc73188c1f2d051d6cfcd06fc920d2a5fa46ba75d10cf1e4
adapter/adapter_config.json	2,155 B	67d422231ccd495178c44f799139b50d7e20a517443a883605f959d432155534

• The GGUF is the adapter merged into the bf16 base, converted with llama.cpp, and quantised to Q4_K_M with an importance matrix computed on wikitext-2's train split (200 chunks). Its tensor data is byte-identical to the file evaluated below (a SHA-256 over all 753 tensors matches); only the metadata was scrubbed for release: general.name set to caveowl-35b-a3b (was a working name), quantize.imatrix.file and quantize.imatrix.dataset removed (they held local paths), and general.sampling.temp set from the base's 1.0 to 0.3, the temperature that was gated.
• The mmproj is stock Ornith's vision projector, unchanged: bartowski's f16 conversion of Ornith-1.5-35B-A3B's vision tower, byte-identical to mmproj-Ornith-1.5-35B-A3B-f16.gguf in bartowski/Ornith-1.5-35B-A3B-GGUF (same sha256). The LoRA does not touch the vision tower, so the stock projector matches the merged model; it is copied here so the repo runs on its own. ornith-ai's own BF16 projector (ornith-ai/Ornith-1.5-35B-A3B-GGUF) should work equally, but was not the one tested.
• The adapter is the final LoRA (r 16, alpha 32, F32 tensors) as saved by Unsloth. Its adapter_config.json differs from the training output only in base_model_name_or_path (the local path replaced by ornith-ai/Ornith-1.5-35B-A3B) and revision (set to the base revision trained on, 10fbf86fed7ecee4a061f8b499a618f46001cac1).

Intended use

Short interactive answers and agent work: chat replies, tool calls, shell and code help, pi-style skills. It was built to cut tokens in an assistant that answers on a phone and in a terminal.

It is not meant for long-form writing. See Limitations.

How to run (llama.cpp)

llama-server \
    --model Ornith-1.5-35B-caveowl-run5-Q4_K_M.gguf \
    --mmproj mmproj-Ornith-1.5-35B-A3B-f16.gguf \
    --jinja --reasoning-format deepseek \
    --temp 0.3 --top-k 20 --top-p 0.95 \
    -ngl 999 --flash-attn on --cache-type-k q8_0 --cache-type-v q8_0 \
    --ctx-size 262144 --parallel 2 \
    --alias caveowl-35b-a3b

• Serve it at temperature 0.3. At 0.6, sampler-check (below) caught it leaving out a max_results the user had asked for in 2 of 90 sampled tool calls; stock Ornith did that 0 times in 360 over four nights. At 0.3 both scored 90/90. The GGUF's embedded general.sampling.temp is 0.3, so llama.cpp's default sampler is the gated one; pass --temp 0.3 explicitly anyway, and set it in your client, because some front ends send their own temperature.
• --jinja is required for tool calling. The chat template embedded in the GGUF is the one the model was trained on: stock Ornith GGUF's (an Unsloth-fixed variant of the Hub template).
• --ctx-size 262144 --parallel 2 gives two slots of 131,072 tokens, which is how it was served during evaluation. Lower it to fit your memory. The Q4_K_M weights need about 22 GB plus the KV cache; on the author's box the whole server sits at about 25 GiB of GPU memory.
• Thinking is on by default. To turn it off per request, send "chat_template_kwargs": {"enable_thinking": false}:

curl -s http://127.0.0.1:8080/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "caveowl-35b-a3b",
  "messages": [{"role": "user", "content": "Why is the sky blue?"}],
  "temperature": 0.3,
  "chat_template_kwargs": {"enable_thinking": false}
}'

The evaluated build was a Vulkan llama.cpp fork (Nathanw1014/llama.cpp, branch strix-halo-vulkan) on an AMD Ryzen AI Max+ 395. Any llama.cpp recent enough to load stock Ornith-1.5-35B-A3B (qwen35moe) should load this GGUF, but caveowl itself was only run on that fork. wikitext-2 perplexity (-c 512, 100 chunks, fork v0.7.8.1) is 7.34 for caveowl against 8.04 for stock Ornith; a lower perplexity on wikitext is not a quality claim, only a check that the merge did no gross damage.

How to use the adapter

The adapter applies to the bf16 base at revision 10fbf86fed7ecee4a061f8b499a618f46001cac1 (Qwen3_5MoeForConditionalGeneration, about 67 GB in bf16). It was trained and saved with Unsloth (2026.10.3) on transformers 5.17.0 and peft 0.21.2. It puts LoRA on the routed experts' fused 3-D gate_up_proj / down_proj parameters through Unsloth's MoE path (target_parameters plus an unsloth_fused_expert_lora layout block in the config), so a plain PEFT install may not load it. Loading it through Unsloth is what was done here:

from unsloth import FastLanguageModel
from peft import PeftModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "ornith-ai/Ornith-1.5-35B-A3B",
    revision="10fbf86fed7ecee4a061f8b499a618f46001cac1",
    max_seq_length=16384, load_in_4bit=False, dtype="bfloat16",
)
model = PeftModel.from_pretrained(model, "stefanj0/caveowl-35b-a3b", subfolder="adapter")

To merge and convert it yourself, merge_and_gguf.sh and merge_owl.py in the caveowl repo are the scripts that produced the GGUF here; they check that untouched tensors, the vision tower and the MTP head come through byte-identical. Only the GGUF has been evaluated; the adapter loaded in transformers has not been gated separately.

Training

Method. bf16 LoRA (not QLoRA), r 16, alpha 32, dropout 0, on all 256 routed experts (gate_up_proj, down_proj), the shared expert, the 10 full-attention layers' q/k/v/o and the 30 Gated DeltaNet layers' in_proj_qkv, in_proj_z and out_proj; about 0.94B trainable parameters. Router, embeddings, lm_head and the vision tower were not trained. One epoch, 241 optimiser steps, batch 1 x 16 gradient accumulation, learning rate 1e-4 cosine with 5% warmup, weight decay 0.01, rows up to 16,384 tokens, seed 42. Every row is rendered through the chat template the server sends, and only assistant turns carry loss. Trained with Unsloth on an AMD Ryzen AI Max+ 395 (gfx1151, ROCm 7.13 PyTorch 2.11) in about 14 hours, peak 84 GiB allocated. Final held-out loss: 1.14 on caveman text pairs, 0.33 on tool rows.

Data. 3,847 rows, 4.8M tokens, 22% of them carrying loss. Shares are relative to the text pairs.

Slice	Share	Rows	What it teaches	Source
Caveman text pairs	base	1,679	the style	prompts from six permissively licensed datasets (below), answers rewritten in caveman style by Claude, through cavegemma's corpus pipeline
Tool calls	25%	439	keep tool calls exact while the prose changes	nvidia/Nemotron-Agentic-v1 (prose rewritten by Claude, calls and results unchanged) and xLAM rows from HuggingFaceTB/smoltalk2
Code anchors	20%	420	stay as good at code as before	stock Ornith's own answers to MBPP and UltraData-Code (sh, R, Rust) tasks, kept only if the dataset's tests pass
Skill trajectories	15% x2	382	open a skill's instructions, run its script, answer from its output	stock Ornith driven through real pi sessions on the author's machine, judged, failures pruned
Agent loops	10%	180	long multi-step tool use with skills listed but not needed	julien-c/synthtraces Qwen3.6 sessions, answers rewritten in caveman style by Claude
Bounded tool calls	25%	560	carry a limit the user stated into the call	stock Ornith's calls on two-requirement prompts, kept only if correct
Thinking-on rows	10%	187	close the reasoning, then answer	stock Ornith's reasoning plus the caveman answer

The text pairs' prompts come from OpenAssistant/oasst2, princeton-nlp/SWE-bench_Verified, ronantakizawa/github-codereview, bigcode/commitpackft (MIT/Apache subset), theblackcat102/evol-codealpaca-v1 and HuggingFaceH4/ultrachat_200k, as selected by cavegemma's build_corpus.py. The skill trajectories cover the skills the author has installed: weather, system health, PDF reading and exploration, summarising, humanising text, explaining simply, and a coffee-price tracker. No training data is published.

Evaluation

Every check ran on stock Ornith-1.5-35B-A3B Q4_K_M and on caveowl run 5 Q4_K_M, one at a time, on the same port, llama.cpp binary and flags, both at temperature 0.3, on 2026-10-16. A check passes when caveowl is at least as good as stock. The tool and code harnesses are public in sjvrensburg/llm-bench-kit; the skill, thinking-on and held-out code gates are in the caveowl repo (eval/, scripts/gate.sh).

Check	What it measures	Stock Ornith @0.3	caveowl run 5 @0.3
sampler-check x30	3 tool prompts x 30 samples at the served temperature; any dropped or invented argument fails the sample	90/90	90/90
tool-harness / pi off / pi shipped	tool-call correctness greedy, through pi's real prompt with thinking off, and with pi's shipped settings	9/9 x3	9/9 x3
code-eval / held-out sh+R+Rust	Python tasks run against tests; 90 held-out UltraData-Code sh, R and Rust tasks never trained on	10/12, 58/90	10/12, 58/90
reasoning leak (thinking off)	16 ordinary chat replies; any <think> text in the visible answer fails; total tokens	0/16, 2,544 tok	0/16, 992 tok
thinking on, empty answers	32 prompts with thinking on; an answer left inside an unclosed think block fails; median answer length	0/32 (312 words)	0/32 (66 words)
skill_gate / + AGENTS.md	pi skill tasks completed (reads the SKILL.md, runs its script, answers from it), without and with the author's agent instructions	44/45, 41/45	45/45, 42/45
skill_gate_wide (8 families)	the same over a wider set of skills and prompts	30/81	38/81

Stock's score on the wide skill gate varied from 28 to 34 of 81 between nights, so differences of a few points there are within noise. All counts are small samples.

Limitations

• Not for long-form prose. In a blind A/B against stock Ornith on 8 writing prompts x 2 samples (podcast segment, essay, lay summary, key-insights list, explainer for a 12-year-old, rewrite, story opening, poem), a Claude judge preferred stock 31 to 1. Expository prose comes out in caveman register and at 30-60% of the requested length (a 300-word explainer came back at 61-88 words). Verse and fiction come out in normal English, but were still judged weaker. A system prompt asking for standard English fixed the register on some tasks and not on others, and did not fix length.
• Terse answers carry less detail. caveowl keeps the facts it states but states fewer of them. Read the answers your use depends on before relying on it.
• Sampler sensitivity. Serve at 0.3; at 0.6 it occasionally drops a requested tool argument (2/90, above). Higher temperatures were not tested.
• Skill use does not generalise far. In earlier runs the model read a skill's instructions reliably only for skill families it had trajectories for. The wide-gate gain is measured on the author's skills and agent, not on yours.
• Code is held, not improved. Scores equal stock; individual tasks can still fail, including with thinking on.
• Speculative decoding. A DFlash drafter for Ornith-35B speeds tool and code turns about 1.5-1.7x on this model but slows short and prose turns, and changes greedy output on longer generations on the tested build. The evaluation above used no drafter.
• Local details in the training data. The skill trajectories and agent-loop rows quote the author's machine: its home directory paths (/home/stefan/...), skill file paths and command output, and the example place names in the weather skill. A light probe of the served model (14 prompts asking for the user's home directory, name, email, phone number, address, suburb, hostname, Signal recipient, coffee purchases, the weather skill's default city, and path completions) reproduced none of these; it answered that it had no access to personal data, or gave generic examples (/home/alice). That probe is not proof that nothing is memorised.
• Evaluated only as the Q4_K_M GGUF on one AMD machine with a Vulkan llama.cpp build. Not tested on CUDA, on other quantisations, or as the bf16 adapter in transformers.
• No safety evaluation beyond the base model's. caveowl is not abliterated; it inherits Ornith's behaviour and refusals as far as the fine-tune left them.

Licences and credits

• This model (adapter and merged GGUF): MIT, as a derivative of Ornith-1.5-35B-A3B under its terms.
• Base model: Ornith-1.5-35B-A3B by ornith-ai, MIT. The vision projector in this repo is ornith-ai's vision tower as converted to GGUF f16 by bartowski, unchanged.
• Pipeline and caveman style: cavegemma and caveman by Julius Brussee, MIT. The caveowl code is MIT (sjvrensburg/caveowl).
• Caveman text pairs: prompts from OpenAssistant/oasst2 (Apache-2.0), princeton-nlp/SWE-bench_Verified, ronantakizawa/github-codereview (permissive subset), bigcode/commitpackft (MIT/Apache subset), theblackcat102/evol-codealpaca-v1 (Apache-2.0) and HuggingFaceH4/ultrachat_200k (MIT); answers rewritten in caveman style by Claude (Anthropic).
• Tool rows (CC-BY-4.0): material derived from Nemotron-Agentic-v1 © NVIDIA Corporation and from xLAM-function-calling-60k © Salesforce (as redistributed in SmolTalk2 by Hugging Face), both licensed under CC BY 4.0. Changes: rows filtered and normalised to one message format, long tool results truncated, Nemotron's assistant prose rewritten in caveman style by Claude; tool calls and results unchanged. A model trained on them must carry this notice.
• Code anchors: prompts and tests from google-research-datasets/mbpp (CC-BY-4.0) and openbmb/UltraData-Code L3 sh/R/Rust (Apache-2.0; its card asks users to respect the source repositories' licences and forbids unchanged redistribution). Answers are the stock model's own; the datasets' reference solutions were never trained on.
• Agent loops: julien-c/synthtraces (MIT), answers rewritten in caveman style by Claude.
• Skill trajectories, bounded tool calls, thinking-on rows: generated locally from stock Ornith.
• Trained with Unsloth; converted, quantised and served with llama.cpp.