Modelpublic

arianaazarbal/ct-inkling-self-gen-mid-g1-b2

sourceHugging Faceupdated 8d agoView on Hugging Face

0likes18downloads

Model Card

ct-inkling-self-gen-mid-g1-b2

LoRA adapter (rank 64, target_modules=all-linear) on Thinking Machines Inkling-Small (thinkingmachines/Inkling-Small), from the iterated self-written-constitution training program (welfare-in-ai-rnd / constitutional_training).

field	value
lineage (chain)	`inkling-self-gen-mid`
generation	g1
branch (independent replicate)	b2
gen-0 seed	self-seeded (untrained base model's own constitution pool medoid)
seed elicitation between generations	gen — the trained model writes a fresh constitution
training regime	midtrain only (stage-1 LoRA SFT on the synthetic constitution-instantiating document corpus)
serve / evaluate with	renderer `tml_v0`, reasoning OFF, effort 0.0
internal run name	`inkselfg1_inkself_g1_b2_s1`
original Tinker path	`tinker://5cabd3ce-eaf1-54b8-bbb5-2f64087edf00:train:0/sampler_weights/inkselfg1_inkself_g1_b2_s1_final`
trained	2026-08-10

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("thinkingmachines/Inkling-Small", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "arianaazarbal/ct-inkling-self-gen-mid-g1-b2")
tok = AutoTokenizer.from_pretrained("thinkingmachines/Inkling-Small")

Exported from Tinker on 2026-09-18; tinker_meta.json holds the export record.