qwen3-0.6b-nhl-polymarket-grpo-v2b API & Inference Endpoint

Result

Walk-forward eval (eval/walkforward_v2b_edgestake.json in sach0312/nhl-polymarket-trading), identical accounting to all prior evals:

Table with columns: split, model stake, edge stake, valid outputs that trade
split	model stake	edge stake	valid outputs that trade
test (227 games)	$0.0000	$0.0000	0 of 855 (all no_trade)
val (621 games)	$0.0000	$0.0000	0 of 2,332 (all no_trade)

Every parseable output was no_trade on both splits; 56–63% of outputs contained no JSON at all. The policy is a degenerate no-trade agent.

History

500/500 steps trained cleanly (~67 min, a10g-small), train loss −0.018, reward mean flat near 0 through late training with frac_reward_zero_std rising to 0.6 — the collapse was visible in-training, as in the v1 GRPO runs.
Four documented GRPO attempts across this project (v1 temp 1.8, v1 temp 3.0, v2b original, v2b rerun) all degenerated the same way.
Conclusion: GRPO with this reward structure on a 0.6B model collapses to no-trade regardless of reward shaping. See the grpo-v2 card for the one positive LLM result (GRPO-v2, edge-staked test +$0.0355) and the dataset README for the full policy ranking (champion: LUT teacher + edge staking).

Training

Trainer: TRL GRPOTrainer, adaptive entropy (target 1.2), 8 generations, temp 1.0, lr 1e-5, 500 steps. Metrics in trackio bucket qwen3-nhl-grpo-v2b-stakereward (Space freeze failed; bucket data exists).

Result

Walk-forward eval (eval/walkforward_v2b_edgestake.json in sach0312/nhl-polymarket-trading), identical accounting to all prior evals:

Table with columns: split, model stake, edge stake, valid outputs that trade
split	model stake	edge stake	valid outputs that trade
test (227 games)	$0.0000	$0.0000	0 of 855 (all no_trade)
val (621 games)	$0.0000	$0.0000	0 of 2,332 (all no_trade)

Every parseable output was no_trade on both splits; 56–63% of outputs contained no JSON at all. The policy is a degenerate no-trade agent.

History

500/500 steps trained cleanly (~67 min, a10g-small), train loss −0.018, reward mean flat near 0 through late training with frac_reward_zero_std rising to 0.6 — the collapse was visible in-training, as in the v1 GRPO runs.
Four documented GRPO attempts across this project (v1 temp 1.8, v1 temp 3.0, v2b original, v2b rerun) all degenerated the same way.
Conclusion: GRPO with this reward structure on a 0.6B model collapses to no-trade regardless of reward shaping. See the grpo-v2 card for the one positive LLM result (GRPO-v2, edge-staked test +$0.0355) and the dataset README for the full policy ranking (champion: LUT teacher + edge staking).

qwen3-0.6b-nhl-polymarket-grpo-v2b

README

Result

History

Training

Explore FriendliAI today

README

Result

History

Training