lfm2.5-1.2b-noval-instruct-paramdelta

LFM2.5-1.2B Noval — Instruct (ParamΔ)

Zero-cost instruct graft on the domain CPT model via ParamΔ: Θ = noval + λ·(Instruct − Base). No training — only weight arithmetic.

Overview

• Stage: ParamΔ instruct graft (no training)
• Lineage: Base → CPT (noval) → ParamΔ (this model)
• Method: ParamΔ [paramdelta], an instance of task arithmetic [taskarith]: the official post-training delta (Instruct − Base) is summed onto the CPT'd weights (λ=1.0), transferring instruction-following while keeping domain knowledge. Element-wise merge in fp32, saved bf16; tokenizer + chat template + generation config from the Instruct model.
• Domain: fictional — Orbital Mining Corporation (OMC) technical docs + Mars Express telemetry.

Training

Evaluation

No task metrics for this checkpoint (intermediate / no-training stage).

Intended use & limitations

Chat / instruction-following over the OMC / Mars-Express domain at zero post-training cost. Baseline for the agentic SFT round.

Limitations:

• ParamΔ is validated in the literature on 8B–70B; on a 1.2B (more entangled weights) transfer quality is less certain — measure, don't assume.
• λ=1.0 not swept; mild domain dilution observed (e.g. bare acronym 'OMC').
• NOT agentic: does not execute tool-calling trajectories (see -agentic).

Innovations tested

• ParamΔ at 1.2B — the paper validates 8B–70B; we test the zero-cost instruct graft at the 1.2B scale [paramdelta].

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("atenareply/lfm2.5-1.2b-noval-instruct-paramdelta")
model = AutoModelForCausalLM.from_pretrained("atenareply/lfm2.5-1.2b-noval-instruct-paramdelta")
msgs = [{"role": "user", "content": "What does the Orbital Mining Corporation do?"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True)
print(tok.decode(model.generate(**inputs, max_new_tokens=256)[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Citations

• ParamΔ — Cao et al., ParamΔ for Direct Weight Mixing, ICLR 2025. arXiv:2504.21023
• Task arithmetic — Ilharco et al., Editing Models with Task Arithmetic, 2023. arXiv:2212.04089

---

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