You can deploy this model and start earning money today!

GoodStartLabs/
qwen3-8b-openspiel-mix8-curriculum-100iter

Name: qwen3-8b-openspiel-mix8-curriculum-100iter
Availability: InStock

peft

safetensors

lora

reinforcement-learning

Views

No views yet

qwen3-8b-openspiel-mix8-curriculum-100iter

LoRA adapter for Qwen/Qwen3-8B trained with Tinker / tinker-cookbook on an 8-game OpenSpiel mix using a 3-phase scripted-opponent curriculum.

Run


Run name	`openspiel-qwen3-8b-mix8-curriculum-100iter-202604260230`
Wandb	`good-start-labs/Neurips 2026` — run name `openspiel-qwen3-8b-mix8-curriculum-100iter-202604260230`
Tinker train id (final)	`884edcc2-4da0-5569-a16f-8f1fec64914e`
Source checkpoint	`tinker://884edcc2-4da0-5569-a16f-8f1fec64914e:train:0/sampler_weights/final`
Date	2026-04-26

Note: checkpoints.jsonl for this run contains two final entries due to a supervisor restart at iter 100; this adapter is from the last entry (884edcc2-4da0-5569-a16f-8f1fec64914e), the run that completed cleanly.

Training config

Base model: Qwen/Qwen3-8B
Renderer: qwen3_disable_thinking (prefills <think></think>)
LoRA: rank 32, alpha 32, targets {q,k,v,o,gate,up,down}_proj + lm_head
Optim: lr 1e-5, importance-sampling loss, KL coef 0
Rollouts: groups_per_batch=8, group_size=4, max_tokens=32, temperature=1.0
Schedule: n_batches=100, eval_every=10, save_every=10 (checkpoints 10/20/.../100/final)
Execution: stream, 2 minibatches, 1 substep
Returns: normalize_returns=True, invalid_action_penalty=-1.0
Self-play: disabled

Game mix and opponent curricula

mixture_mode="round_robin", seat_mode="uniform", transform_simultaneous_to_turn_based=True. Multi-player games use a 3-phase curriculum keyed on training_fraction = batch_idx / (n_batches - 1), with phase boundaries at 1/3 and 2/3.

Game	Players	Curriculum (easy / medium / hard)
`blackjack`	1p + chance	—
`morpion_solitaire`	1p combinatorial	—
`nim` (misere, piles 1/3/5/7)	2p	random / `nim_greedy` / `nim_epsilon_0.1`
`hanabi`	2p coop, partial info	random / `hanabi_rule_bot` / `hanabi_rule_bot`
`sheriff`	2p bargaining	random / `sheriff_simple` / `sheriff_simple`
`connect_four`	2p zero-sum	random / `one_step_win` / `win_or_block`
`breakthrough`	2p zero-sum	random / `one_step_win` / `win_or_block`
`pentago`	2p zero-sum	random / `one_step_win` / `one_step_win`

System prompt constrains output to ACTION: <index>.

Relation to sibling run

GoodStartLabs/qwen3-8b-openspiel-mix8-curriculum3-100iter uses the same config with group_size=8 instead of 4 (i.e. 2x rollouts per batch). This run is the lower-rollout variant.

Loading

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype="bfloat16", device_map="auto")
tok  = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
model = PeftModel.from_pretrained(base, "GoodStartLabs/qwen3-8b-openspiel-mix8-curriculum-100iter")