SkyRL recipes: full convergence demonstrated only on GSM8K — capture 8B LoongRL/BIRD curves

[sfc-gh-kganesan]

Summary

Full convergence is demonstrated only for the 0.6B simple_gsm8k recipe. The
8B long_context_qa (LoongRL) and txt2sql (BIRD) recipes have so far only
been run for short windows (≤10 steps) showing healthy, monotonic learning,
not full convergence curves.

Validated so far

• simple_gsm8k (Qwen3-0.6B): eval pass@1 0.29 → 0.64, reward 0.28 → 0.80
  over 70 steps. ✅ Full convergence.
• long_context_qa 8B (LoongRL): reward climbing over 9 steps. ✅ Learning,
  ⏳ not a full curve.
• txt2sql 8B (BIRD): reward climbing over 10 steps. ✅ Learning, ⏳ not a
  full curve.

Acceptance

• Run LoongRL 8B long enough to publish a convergence curve.
• Run BIRD 8B long enough to publish a convergence curve.
• Attach W&B curves / final metrics to PR recipes/rl/skyrl: Arctic RL × SkyRL recipes (simple / txt2sql / long_context_qa) #32 (or the recipe READMEs).