Modelpublic

maxonxie/nanojev-hard-lr1e5-repro

sourceHugging Facemitupdated 6d agoView on Hugging Face

1likes15downloads

Model Card

nanojev-hard-lr1e5-repro

Personal reproducibility run of NanoJev hard CE SFT (hard_lr1e5, seed 17, 600 steps), starting from the public initialization in C-Tianyu/NanoJev (training_initialization).

Layout mirrors the root checkpoint directory of C-Tianyu/NanoJev (weights + tokenizer + eval artifacts), so you can download this folder and continue training or serve with NanoJev tooling.

This is not the official unified-games-v1 release. Checkpoint selection used min dev `selection_ce` → best_step = 300.

Files (same style as official unified root)

Path	Role
`best.safetensors`	Selected weights (step 300)
`config.json`	Config from this run (may contain original machine paths)
`training_config.json`	Same hyperparams with local absolute paths redacted
`tokenizer/`	Tokenizer
`backbone_config/`	Backbone config
`summary.json`	Metrics + `weights_sha256`
`train_log.json`	Step log
`predictions_{dev,test,ood,calibration}.jsonl`	Split predictions
`initial_dev.jsonl` / `initial_dev_metrics.json`	Init-time eval dump
`target_audit.json`	Target audit

Recipe

—Backbone: Qwen/Qwen3-0.6B + set_head=attention
—Init SHA256: 38116340795de1c82369b7fe15819d92d79600a7b4dc7a3cd0d4390cb6782639
—Data: NanoJev unified hard mix
—Loss: hard CE; backbone lr 1e-5, head lr 1e-4; seed 17; 600 steps; eval every 100
—Selected weights SHA256: f53eadbd34eb5a1040d9d579c64bebde45a7180fa3b012d187e64c118c09022f
—best_dev_selection_ce ≈ 0.636 · test ≈ 0.678 · ood ≈ 0.759

Download → continue train / serve

bash

huggingface-cli download maxonxie/nanojev-hard-lr1e5-repro \
  --local-dir ./nanojev-hard-lr1e5-repro

# serve
python -m research.toy.serve_decisions \
  --checkpoint-dir ./nanojev-hard-lr1e5-repro \
  --host 127.0.0.1 --port 8765

# continue SFT: point --init-checkpoint (or your NanoJev flag) at this directory
# and set --input to your hard data path on the new machine.

Exact CLI flags follow your NanoJev checkout; the checkpoint directory shape matches what serve_decisions / train scripts expect for a finished run.

License / attribution

—NanoJev / OpenJev (MIT)
—Backbone: Qwen/Qwen3-0.6B
—Init & reference: C-Tianyu/NanoJev
—Study notes: OpenAGI-Go/agimind → notes/nanojev-2026-09/