JMaxCool/bookend-KING-CXXVI-10kDPO-FINAL

bookend-v125-step80

Merged BF16 weights from bookend DPO on the v125 king pool, exported at global training step 80 (SAVE_STEPS=10 → checkpoint-80 merged with LoRA).

Use this directory as a vLLM / local duel / private submit model path.

Contents

Item	Role
model-*.safetensors + model.safetensors.index.json	Sharded merged Qwen3.6-35B-A3B MoE
config.json, tokenizer.json, chat_template.jinja	HF layout for inference
train_bookend_v125_step80.py	Training recipe (check-setup / train / export / show-recipe)

Training data: data/dpo/bookend_dpo_pairs.jsonl (~3277 pairs; join from data/dpo/bookend_dpo_parts/ if missing). See docs/training/BOOKEND_DPO_TRAINING.md.

Base model for training: local_king/king_cxxiv (or current reign king path via KING=...).

Quick start — reproduce training

From repo root (8× GPU recommended):

set -a && source .env && set +a
pip install -e '.[train]'   # once

# merged pairs file
python scripts/train/split_train_pairs.py --join \
  --src data/dpo/bookend_dpo_pairs.jsonl \
  --parts-dir data/dpo/bookend_dpo_parts

python checkpoints/bookend-v125-step80/train_bookend_v125_step80.py check-setup
python checkpoints/bookend-v125-step80/train_bookend_v125_step80.py train

View baked-in hyperparameters:

python checkpoints/bookend-v125-step80/train_bookend_v125_step80.py show-recipe

Defaults in BookendV125Step80Recipe (override with env):

Variable	Default
KING	local_king/king_cxxiv
TRAIN_PAIRS	data/dpo/bookend_dpo_pairs.jsonl
OUT_DIR	checkpoints/bookend_v125_lora
NUM_GPUS	8
EPOCHS	2
LR / BETA	5e-7 / 0.1
BATCH_SIZE / GRAD_ACCUM	1 / 8
SAVE_STEPS	10
TARGET_STEP	80

Training runs export automatically after DPO (merge checkpoint-80 into this folder).

Re-export only (LoRA run already exists):

OUT_DIR=checkpoints/bookend_v125_lora TARGET_STEP=80 \
  python checkpoints/bookend-v125-step80/train_bookend_v125_step80.py export

Inference (local eval)

vllm serve checkpoints/bookend-v125-step80 \
  --tokenizer assets/tokenizers/Qwen3.6-35B-A3B \
  --port 8000

python scripts/local_eval/run_duel.py --mode duel \
  --candidate-base-url http://localhost:8000/v1 \
  --candidate-model checkpoints/bookend-v125-step80 \
  --n-samples 20

Run preflight_gate.py before any on-chain private submit.

Private submit (example)

set -a && source .env && set +a
.venv/bin/albedo submit-private \
  --path checkpoints/bookend-v125-step80 \
  --name bookend-v125-step80 \
  --coldkey JMax-1 \
  --hotkey sn97-15 \
  --yes

Ensure the hotkey wallet matches the registered UID (see project field notes if Custom error: 19).

Related checkpoints

Sibling exports in checkpoints/: bookend-v125-step50, step60, step70, same training run at different checkpoint-* steps.

Notes

• Validators expect BF16 merged weights, not raw LoRA adapters.
• Bookend pairs target submit-aligned chosen tails; monitor simple_bot / pre-eval on dashboard after submit.
• Do not commit secrets; use repo root .env for HF_TOKEN and wallet passwords.