lfm2.5-1.2b-asterion-agentic

LFM2.5-1.2B Asterion — Agentic (SFT v2)

Verifiable satellite-operations assistant: telemetry alert triage + native tool-calling + raw-log extraction. LoRA SFT on the ParamΔ model with the v2 confound-free dataset.

Overview

• Stage: Agentic SFT v2 (LoRA, merged)
• Lineage: Base → CPT (asterion) → ParamΔ (instruct) → Agentic SFT v2 (this model)
• Method: Supervised fine-tuning (LoRA r=32/α=64, all-linear; merged) [lora] on the v2 dataset: raw-vs-clean tool output decorrelated from difficulty (independent coin flip 0.5 — kills the v1 'raw ⇒ genuine' shortcut), Track C multi-QA, teacher reasoning re-wrapped as <think> with assistant-only loss [apigen]; ≥50% general blend against collapse [collapse].
• Domain: fictional — Orbital Mining Corporation (OMC) technical docs + 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 for the Asterion fleet: classify an ALERT (routine vs genuine), run the maintenance loop on genuine anomalies, extract values from raw logs. Use INSIDE a tool-execution loop.

Limitations:

• The verifiable task is SATURATED (pass@1 0.986 > the 0.85 RL gate) — no RL round was run on this model by design (RLVR needs headroom).
• Fictional domain; the genuine class is largely synthetic.
• Tool-calling REQUIRES a multi-turn loop that executes tools and feeds results back.

Innovations tested

• Confound-free v2 data — raw/clean tool-output format is an independent per-task coin flip, decoupled from difficulty and 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/lfm2.5-1.2b-asterion-agentic")
model = AutoModelForCausalLM.from_pretrained("atenareply/lfm2.5-1.2b-asterion-agentic")
TOOLS = [...]  # the 8 OpenAI-style tool schemas (see noval-corp/config/curation.py: TOOL_SCHEMA)
messages = [{"role": "system", "content": "You are an OMC 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_start|>[...]<|tool_call_end|>, execute, append a {"role":"tool",...} message, repeat.

Reference loop: noval-corp/scripts/eval_agentic.py::run_episode.

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
• Model collapse (≥50% real data) — Shumailov et al., 2024. arXiv:2404.01413

---

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