rgh — Epago SN36 private backup

Training artifacts for the SN36 challenger built from Alibaba-NLP/Tongyi-DeepResearch-30B-A3B @ hf:4b0ac5767427a55d08a254f0367e2934976598e0.

Layout

Path	What
challenger_v1/ v2/ v3/	merged 30B weights (17 shards each)
adapter_v1/ … adapter_v6b/	PEFT LoRAs (r=32, alpha=64, q/k/v/o_proj)
sft_stage/	SFT mixes including train_v3.jsonl … train_v6b.jsonl
dataset/train_final.jsonl	v1 SFT mix (913 episodes)
data/	harvests, teacher rollouts, evals, bakeoffs
splits/	train/eval task splits
practice-tasks/	minted POOL1 / SCI4 practice tasks
practice-v1/	entities, manifest, and corpus.db if present
minerlab/	local train/eval/harvest scripts and analyses
notes/six_research_episodes.txt	episode writeup

Not in this repo: secrets, wallets, upload-auth*, the public Tongyi king snapshot. Merged challenger_v4+ weights are rebuilt from challenger_v3 + the matching adapter.

Load a merged challenger

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("danyloooah/rgh", subfolder="challenger_v3")
model = AutoModelForCausalLM.from_pretrained(
    "danyloooah/rgh", subfolder="challenger_v3", torch_dtype="auto", device_map="auto"
)

Load a LoRA

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Alibaba-NLP/Tongyi-DeepResearch-30B-A3B"
tok = AutoTokenizer.from_pretrained("danyloooah/rgh", subfolder="adapter_v6b")
base = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "danyloooah/rgh", subfolder="adapter_v6b")