atenareply

gemma-4-12b-asterion-instruct-paramdelta

README

License: apache-2.0

Overview

Stage: ParamΔ instruct graft (no training)
Lineage: gemma-4-12B → CPT (asterion) → ParamΔ (this model)
Method: ParamΔ [paramdelta] / task arithmetic [taskarith]: the official post-training delta (gemma-4-12B-it − gemma-4-12B) — which carries chat AND the native tool-calling format — is summed onto the CPT'd weights (λ=1.0). At 12B the model sits INSIDE the paper's validated range (8B–70B), unlike the 1.2B round. Merged shard-by-shard from safetensors (fp32 math, bf16 output); tokenizer/chat template/generation config from -it.
Domain: fictional — Asterion Space Operations (24-satellite constellation: alert triage, anomaly investigation, subsystem ops) + reused Mars Express telemetry.

Training

Table

Method	weight arithmetic only — NO gradient training
λ (instruct delta scale)	1.0
Merge	shard-streaming (peak RAM ≈ one shard × 3, not 3 full models)

Evaluation

Table with columns: Metric, Value, Note
Metric	Value	Note
PPL Asterion held-out (informative only)	13.68	CPT parent: 1.83; plain -it: 64.28 — the graft carries the domain delta but raw PPL is template-bound-model artifact
PPL general (informative only)	162.8	plain -it: 302.97 — see limitations
chat smoke EN/IT (human gate)	PASS	coherent, domain-anchored, both languages
byte-level merge verification	PASS	sha256-verified upload; CPT delta present on 119,980/200,000 sampled weights

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

Chat / instruction-following / native tool-calling over the Asterion domain at zero post-training cost. Starting checkpoint for the agentic SFT.

Limitations:

λ=1.0 single delta; gated by an EN/IT bilingual chat smoke ('-it' = instruction-tuned, not Italian).
Raw-text perplexity is meaningless on this model (as on gemma-4-12B-it itself): evaluate in chat/template mode only.
Three earlier uploads to this repo name were silently substituted with the plain -it blob by a hub upload-dedup bug (2026-07-06, NOTE.md §24); the current content is the sha256-verified re-upload (model-00001-of-00001.safetensors, sha f251c432…).
NOT agentic-specialized (see -agentic).

Innovations tested

Shard-streaming ParamΔ — the 12B merge never materialises full fp32 state dicts (~144 GB) — tensors are read lazily per shard and written bf16 [paramdelta].

Usage

python
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("atenareply/gemma-4-12b-asterion-instruct-paramdelta")
model = AutoModelForCausalLM.from_pretrained("atenareply/gemma-4-12b-asterion-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).

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

atenareply

Model Tree

Base

atenareply/gemma-4-12b-asterion

Fine-tuned

this model

Input Modalities

TextAudioImageVideo

Output Modalities

Text

Supported Functionality