Gemma-4-26B-A4B agentic LoRA — run4 (OpenCode format)

A small LoRA for google/gemma-4-26B-A4B-it that nudges it toward agentic coding, trained on coding-agent trajectories rewritten into OpenCode's tool vocabulary, system prompt and tool-result formats. Applied at low strength with thinking off, it keeps the model's score on an agentic coding benchmark.

Sibling: run2 — the same trajectories in their original OpenHands format. run2 measured better, even inside OpenCode, and is the one to try first; run4 is published as the OpenCode-format counterpart.

Files

Use it with llama.cpp — at strength 0.15–0.25, thinking off

llama-server -m gemma-4-26B-A4B-it-Q8_0.gguf \
  --lora-scaled gemma4-agentic-run4-tok5M-lora-bf16.gguf:0.25 \
  --jinja --chat-template-file chat_template.jinja --reasoning off \
  --ctx-size 131072 --parallel 1 -ngl 999 \
  --temp 1.0 --top-p 0.95 --top-k 64

• Keep the strength low. Use --lora-scaled …:0.15 to :0.25, not --lora (which applies 1.0). It degrades quickly above that: 40/42 at 0.5, 26/42 at 0.75.
• Thinking off. It was trained and measured with thinking disabled; with thinking on it measured the same as stock Gemma (no benefit, no harm).
• chat_template.jinja is the one shipped in the base instruct repo. Any GGUF of the instruct model works; the numbers below are on Q8_0.

With transformers/PEFT: PeftModel.from_pretrained(base, "Myric/gemma4-26b-a4b-agentic-lora-run4", subfolder="tok5M"), then scale the adapter (or merge at 0.25). The config's base_model_name_or_path is a local path — pass the base model yourself.

Measured

42-task agentic coding suite (a real OpenCode tool loop, graded by hidden pytest), thinking off, DGX Spark:

At 0.15–0.25 it matches stock's score; its token counts sit inside stock's range, so unlike run2 there is no clear token saving. Token counts vary a lot between identical runs (the two 0.15 runs differ by 2x), so treat single-run figures as indicative.

Training

• Adapted: attention q/k/v/o_proj and each layer's dense MLP gate/up/down_proj; LoRA rank 32, alpha 64, no dropout (37.2 M parameters). The 128 routed experts are frozen.
• Data: ~5 M tokens (89 steps) of a mix: 45% agentic — resolved nebius/SWE-rebench-openhands-trajectories, windowed to ~10k tokens and re-expressed in OpenCode 1.18.31's tools, system prompt and result formats; 35% general chat and 20% tool calls from HuggingFaceTB/smoltalk + Team-ACE/ToolACE, with refusals filtered out.
• Loss only on tokens the model generates, rendered with the instruct repo's own chat template and tool tokens.
• Trained on one DGX Spark (GB10).

License

Use is subject to the Gemma Terms of Use. Training data: SWE-rebench OpenHands trajectories (CC-BY-4.0, Nebius), SmolTalk (Apache-2.0, Hugging Face), ToolACE (Apache-2.0).