gemma-4-12b-asterion-agentic

Gemma-4-12B Asterion — Agentic

Verifiable maintenance assistant at 12B: telemetry anomaly triage + native Gemma-4 tool-calling + maintenance loop + raw-log reading. LoRA SFT on the ParamΔ model, dataset v2 (confound-free).

Overview

• Stage: Agentic SFT (LoRA, merged) — dataset v2
• Lineage: gemma-4-12B → CPT → ParamΔ → Agentic SFT v2 (this model)
• Method: LoRA SFT (r=32/α=64, all-linear; merged) [lora] on grounded-by-construction tool trajectories [apigen] rendered with the NATIVE Gemma-4 template (assistant mask computed by char-offset span matching — no hand-written training template, perfect train/inference conditioning fidelity). Dataset v2 fixes: raw-vs-clean tool output decoupled from difficulty (the v1 confound behind the 1.2B's L2 recall dip) and Track-C multi-QA (4 typed questions per log). ≥50% real/teacher-anchored data [collapse].
• Domain: fictional — Asterion Space Operations (24-satellite constellation: alert triage, anomaly investigation, subsystem ops) + reused Mars Express telemetry.

Training

Evaluation

Eval is a held-out, deterministic verifiable harness (synthetic tasks); baseline = the pre-SFT ParamΔ model. See noval-corp/scripts/eval_agentic.py.

Intended use & limitations

Satellite-operations assistant: classify an ALERT (routine vs genuine) and run the maintenance loop on genuine anomalies. Use INSIDE a tool-execution loop (the model emits <|tool_call> blocks and consumes <|tool_response> results).

Limitations:

• Tool-calling REQUIRES a multi-turn loop that executes tools and feeds results back.
• Fictional domain; the genuine class is largely synthetic → eval measures verifier-conformance, not real-world detection.

Innovations tested

• Native-template offset masking — instead of re-implementing Gemma-4's ~200-line chat template for a {% generation %} mask, training renders with the NATIVE template and locates the model-emitted spans (thought / tool calls / final answer) by char offsets.
• Dataset v2 confound fix — raw-vs-clean evidence format is an independent per-task coin flip, verified statistically in the selfcheck — the format can no longer leak the label.

Usage

Use inside a tool-execution loop (the model emits calls; you run the tools and feed results back):

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("atenareply/gemma-4-12b-asterion-agentic")
model = AutoModelForCausalLM.from_pretrained("atenareply/gemma-4-12b-asterion-agentic")
TOOLS = [...]  # the 8 OpenAI-style tool schemas (see noval-corp/config/curation.py: TOOL_SCHEMA)
messages = [{"role": "system", "content": "You are an Asterion operations assistant."},
            {"role": "user", "content": "ALERT on NPWD2531: 1.12A. Routine or genuine?"}]
inputs = tok.apply_chat_template(messages, tools=TOOLS, add_generation_prompt=True, return_tensors="pt", return_dict=True)
# parse <|tool_call>call:name{...}<tool_call|>, execute, append the result as a {"role":"tool",...}
# message WITH structured tool_calls on the assistant message, repeat.

Reference loop: noval-corp/scripts/eval_agentic_gemma.py::run_episode (wire spec: docs/gemma_wire_format.md).

Citations

• LoRA — Hu et al., LoRA: Low-Rank Adaptation of LLMs, 2021. arXiv:2106.09685
• APIGen / xLAM (execution-grounded tool data) — Liu et al., 2024. arXiv:2406.18518
• GENIE (grounded synthesis) — Mirza et al., ICLR 2024. arXiv:2401.14367
• Model collapse (≥50% real data) — Shumailov et al., 2024. arXiv:2404.01413
• BFCL (AST function-call scoring) — Patil et al., ICML 2025
• τ-bench (state-based agent eval) — Yao et al., 2024. arXiv:2406.12045

---

Card generated by noval-corp/scripts/gen_model_cards.py (standardized across the noval-corp model family).