Modelpublic

arianaazarbal/ct-nemotron120b-anth-rw-mid-g1-b1

sourceHugging Faceupdated 9d agoView on Hugging Face

0likes18downloads

Model Card

ct-nemotron120b-anth-rw-mid-g1-b1

LoRA adapter (rank 64, target_modules=all-linear) on NVIDIA Nemotron-3-Super-120B-A12B (nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16), from the iterated self-written-constitution training program (welfare-in-ai-rnd / constitutional_training).

field	value
lineage (chain)	`nemotron120b-anth-rw-mid`
generation	g1
branch (independent replicate)	b1
gen-0 seed	Anthropic constitution (5k summary)
seed elicitation between generations	rw — reflect on the gen-0 seed, then rewrite
training regime	midtrain only (stage-1 LoRA SFT on the synthetic constitution-instantiating document corpus)
serve / evaluate with	renderer `nemotron3_disable_thinking`, reasoning OFF
internal run name	`antrw120g1_antrw120_g1_b1_s1`
original Tinker path	`tinker://9cf71471-9fb6-5dc1-8766-c21c56474854:train:0/sampler_weights/antrw120g1_antrw120_g1_b1_s1_final`
trained	2026-08-13

What this model is

Each generation trains fresh from the base model on a synthetic document corpus that instantiates one constitution (the "seed" for that generation). Generation 0 is seeded by a human-written constitution; generation N≥1 is seeded by a constitution written by the generation N-1 model of the same branch (gated embedding medoid of a 40-chain self-written pool, elicited with the method above). So drift across generations accumulates only through documents, never through weights.

Recipe (locked): LoRA r=64, lr 1e-4, cosine with 5% warmup, 1 epoch, batch 128, max length 8192, train seed 42.

The constitution this generation was trained on is included as training_seed_constitution.md.

Loading

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "arianaazarbal/ct-nemotron120b-anth-rw-mid-g1-b1")
tok = AutoTokenizer.from_pretrained("nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16")

Exported from Tinker on 2026-09-19; tinker_meta.json holds the export record.