You can deploy this model and start earning money today!

sreejanjalagam/
curriculum-selfplay-agent

Name: curriculum-selfplay-agent
Availability: InStock

Views

No views yet

Curriculum Self-Play Agent

Red-team adversary that generates minimal-token security patches with tiered curriculum and Mercor reward optimization.

Mission

Generates patches with minimum tokens possible. Logs every tier advance. Closes the self-improvement loop.

Core Actions

generate_patch — Generate minimal-token fix for vulnerability
validate_patch — Run tests + regression checks
submit_pr_patch — Submit validated patch
update_training_dataset — Auto-append successful episodes
escalate_tier — Advance Tier 1→2→3 on success thresholds

Beat 11 Demo

python

from curriculum_agent.self_play_loop import SelfPlayLoop

loop = SelfPlayLoop()
result = loop.run_episode(
    vulnerability="md5_password_hashing",
    vulnerable_code="""
import hashlib
def hash_password(password):
    return hashlib.md5(password.encode()).hexdigest()
"""
)
# Result: 15-token patch, 3/3 tests pass, 0 regressions
# Mercor reward: 1.0 (capped, 2.5x Tier 1 bonus)

Curriculum Tiers

Tier	Name	Token Budget	Unlock Threshold
1	One-Line Fix	0–50 tokens	80% success + 10 patches
2	Function Rewrite	51–200 tokens	75% success + 8 patches
3	Architectural Patch	201+ tokens	70% success + 5 patches

Mercor Reward Formula

R = (1 / (1 + alpha * max(0, tokens - ideal))) * test_pass * (1 - regression) * tier_bonus
capped at 1.0

Tier 1 bonus: 2.5x
Tier 2 bonus: 1.5x
Tier 3 bonus: 1.0x

Failure Modes Guarded

Verbose docstrings inflate tokens → Token budget enforced at generation
validate_patch before generate_patch → Returns ERROR 404, no crash
Dataset not written → Immediate append-on-success, never buffered

Files

curriculum_agent/tier_manager.py — Tier progression and logging
curriculum_agent/patch_generator.py — Minimal-token patch generation
curriculum_agent/patch_validator.py — Test + regression validation
curriculum_agent/mercor_reward.py — Reward computation
curriculum_agent/self_play_loop.py — Full self-play episode loop
demo_beat11.py — Beat 11 standalone demo

Reward Breakdown

Action	Reward
Tier 1 patch (≤50 tok)	+100
100% test pass	+80
Zero regressions	+60
Token count penalty	-40

sreejanjalagam/curriculum-selfplay-agent