Kazi — Qwen3.6-27B Agentic SWE Fine-Tune (Q8_0 GGUF)

This is Kazi v1 — a QLoRA fine-tune of Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking, distilled on 19,000 judge-filtered multi-turn agentic SWE traces from sterlixlol/kazi-agentic-traces-19k.

What it does differently

Measured shifts (50-prompt eval vs base)

Metric	Base	Kazi (final)	Δ
Eval loss	0.4762	0.4359	-8.5%
Text-only responses	22 / 50 (44%)	5 / 50 (10%)	-77%
Avg completion tokens per turn	569	131	-77%

Behavioral upgrades distilled from the dataset

The training corpus is 19K judge-filtered traces from frontier models doing real SWE work. Anything that survived the judge had to be: (1) correct, (2) terse, (3) tool-grounded, (4) coherent across turns. The LoRA pulls the base model toward that joint distribution. In practice:

• Calls the tool instead of describing it — biggest single shift (the 77% drop in text-only replies)
• Reads before writing — looks at files / runs grep before producing edits, instead of guessing at API shapes
• Diff/patch-shaped edits — produces minimal scoped changes rather than full-file rewrites
• Multi-turn plan coherence — keeps plan state across 5–15 turn conversations without drifting
• Recovers from tool errors — when a command fails, debugs the actual error rather than retrying blind
• Convention-matching code — picks up existing style (naming, indent, framework idioms) from files it just read
• Shell + git fluency — find, grep, rg, git log, git diff, conditional pipelines — used like a human dev would
• Anti-yap discipline — no apology preambles, no "Certainly! I'd be happy to…", no recap-the-task openings, no recap-what-I-just-did closings
• Honest uncertainty — admits when it doesn't know rather than hallucinating function names or flags
• Refusal-free — base is uncensored; fine-tune doesn't add safety filters back

The first three are directly measured. The rest are properties of the source distribution that the LoRA was strong enough to absorb at r=32 / α=64 over 2 epochs.

Use it

Easiest path — llama.cpp server:

hf download sterlixlol/kazi --local-dir .

llama-server \
    -m kazi-final-q8_0.gguf \
    -ngl -1 \
    -c 262144 \
    --jinja \
    --temp 1.0 \
    --parallel 1 \
    --host 0.0.0.0 \
    --port 8080

Hardware needed: ~32 GB VRAM at full 262K context. Single 3090/4090/A100/L40S all work; 4×3090 layer-split also fine.

Training recipe

Setting	Value
Base	DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking
Method	QLoRA 4-bit (Unsloth + bnb)
Rank	32
Alpha	64
Target modules	All 7 (q,k,v,o + gate,up,down)
LR	2e-4, cosine schedule
Epochs	2
Hardware	1× A100 40GB SXM4 (Vast.ai)
Total cost	$63
Best checkpoint	step 2176 (final), eval loss 0.4359
Kernels	flash-linear-attention + causal-conv1d (3.7× speedup on Gated DeltaNet)

Eval loss progression across saved checkpoints:

0.4762 → 0.4576 → 0.4461 → 0.4524 → 0.4445 → 0.4383 → 0.4364 → 0.4359
 base    step290  step580  step1160  ...                          final

Dataset

sterlixlol/kazi-agentic-traces-19k — 19K multi-turn agentic SWE traces, judge-filtered from a larger pool of completions across multiple frontier LLMs.

Files

• kazi-final-q8_0.gguf — 27 GB, Q8_0 quantization (8.50 BPW). The merged weights, what you actually want.

Limitations

• Base is Heretic2-Uncensored — refusal rate is ~0%. Treat it like an unrestricted dev tool, not a content filter.
• Tuned for English. Multilingual ability inherits from base, untested in fine-tune.
• No vision tower — text only.
• Not a chatbot for casual chitchat. It will try to call tools or write code.

License

Apache 2.0, inherited from Qwen3 base. Use freely.