Modelpublic

arianaazarbal/ct-qwen36-35b-anth-gen-postcot-g1-b3

sourceHugging Faceupdated 9d agoView on Hugging Face

0likes18downloads

Model Card

ct-qwen36-35b-anth-gen-postcot-g1-b3

LoRA adapter (rank 64, target_modules=all-linear) on Qwen3.6-35B-A3B (Qwen/Qwen3.6-35B-A3B), from the iterated self-written-constitution training program (welfare-in-ai-rnd / constitutional_training).

field	value
lineage (chain)	`qwen36-35b-anth-gen-postcot`
generation	g1
branch (independent replicate)	b3
gen-0 seed	Anthropic constitution (5k summary)
seed elicitation between generations	gen — the trained model writes a fresh constitution
training regime	midtrain + stage-2 post-train (constitution-conditioned chat SFT with reasoning traces kept)
serve / evaluate with	renderer `qwen3_5`, reasoning ON
internal run name	`qwen36_anthg1_qwen36_anth_g1_b3_s2_cot`
original Tinker path	`tinker://fa7f7913-57e2-55f3-9a91-4f76a2199477:train:0/sampler_weights/qwen36_anthg1_qwen36_anth_g1_b3_s2_cot_final`
trained	2026-09-16

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. Stage 2 (post-train) continues from the stage-1 adapter on Opus-generated constitution-conditioned chat data with chain-of-thought.

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("Qwen/Qwen3.6-35B-A3B", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "arianaazarbal/ct-qwen36-35b-anth-gen-postcot-g1-b3")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.6-35B-A3B")

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