lfm2.5-1.2b-noval-agentic

LFM2.5-1.2B Noval — Agentic

Narrow, verifiable maintenance assistant: telemetry anomaly triage + native tool-calling + autonomous maintenance loop + raw-log reading. LoRA SFT on the ParamΔ model.

Overview

• Stage: Agentic SFT (LoRA, merged)
• Lineage: Base → CPT (noval) → ParamΔ (instruct) → Agentic SFT (this model)
• Method: Supervised fine-tuning (LoRA r=32/α=64, all-linear; merged) [lora] on grounded tool-use trajectories. Tool arguments are supplied by code (exact, label-correct) and the teacher writes only the reasoning/prose — grounded-by-construction [apigen]; ≥50% real data to avoid collapse [collapse]. Native LFM2 tool format (<|tool_call_start|>[...]<|tool_call_end|>); assistant-only loss; best-checkpoint by eval_loss.
• 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

Spacecraft-operations assistant for OMC/Mars-Express telemetry: classify an ALERT (routine vs genuine), and on a genuine anomaly run the maintenance loop (consult SOP → issue command → schedule maintenance → open incident). Use INSIDE a tool-execution loop that returns tool results (the model emits calls and consumes results — it is the decision brain, not the tools).

Limitations:

• Tool-calling REQUIRES a multi-turn loop that executes tools and feeds results back; in plain chat (no tool execution) it emits call markers and stalls/hallucinates.
• Recall on genuine anomalies is 0.63 (≈37% missed) — conservative (precision 1.0). Misses concentrate in L2 (clean-count high-persistence cases); a raw-vs-clean tool-format confound in the v1 data is the suspected cause (fixed in dataset v2).
• Fictional domain; the genuine class is largely synthetic (one real episode) → eval measures conformance to the verifier, not real-world detection.

Innovations tested

• Verifier-as-reward — a deterministic verifier (BFCL-style AST scoring [bfcl] + state-based action check [taubench]) doubles as both the eval and the RFT reward.
• Difficulty-graded agentic curriculum — tasks tagged L1–L4; hard tiers hand the model RAW telemetry rows to read+count (vs a clean aggregate) — log-reading inside the loop.
• Grounded-by-construction trajectories — code supplies exact tool args/labels, the teacher writes only prose [apigen][genie] — near-zero tool-arg hallucination.
• Portable dataset — stored as model-agnostic structured messages_json → the SAME dataset trains LFM2-1.2B and (next) Gemma-4-12B, each via its own chat template.

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-noval-agentic")
model = AutoModelForCausalLM.from_pretrained("atenareply/lfm2.5-1.2b-noval-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
• 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
• RFT / rejection sampling — Yuan et al., 2023. arXiv:2308.01825

---

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