I am building BedirAI (Apache-2.0, https://github.com/bedirai/bedir-ai), an open-source benchmark and verifier for adversarial LLM debate. Instead of letting models win debates with fluent rhetoric or invented quotes, every quote in a debate is checked byte-for-byte with SHA-256 against the source text before a judge scores it.
What are this project's goals? How will you achieve them?
1. Grow our EN/TR/DE pilot dataset (https://bedirsavasi.com/data/pilot-benchmark.json) from the initial verified set to 1,000+ multi-document claims.
2. Run verl (FSDP + vLLM) GRPO self-play ablations on 8B open weights to co-train an evidence-bound debater (Hilal-8B) and a Brier-calibrated judge (Mizan-8B).
3. Publish all code, benchmark splits, calibration scripts, and 8B checkpoints publicly under Apache-2.0, plus host live comparisons on https://bedirsavasi.com/arena/.
How will you know it worked?
Clear numeric bar: Mizan-8B beats uncalibrated LLM-as-a-judge baselines by at least 8 percentage points in Brier score on held-out persuasion and quote-cherry-picking tests, with 0% hash verification errors and reproducible public logs.
How will this funding be used?
- $18,000: on-demand H100/H200 GPU compute (~7,000 GPU-hours) for 8B GRPO runs and Brier calibration sweeps.
- $4,000: frontier evaluation API costs strictly for blind leaderboard baselines (zero training use, enforced in bedir/lineage.py).
At the $5,000 minimum, I would fund the first ~1,800 H100 GPU-hours for 8B pilot ablations and release Mizan-Bench v1.
Who is on your team? What's your track record on similar projects?
Solo founder and engineer Mevlana Yalcin (Izmir, Turkey; https://www.linkedin.com/in/mevlanayalcin/). Already shipped BedirAI v0.3.0 (47/47 CI tests passing), 7 live multi-turn debates at https://bedirsavasi.com/arena/, pilot results (56/56 exact SHA-256 quote matches, 0 failures, 100% directional agreement, 10.0 pp mean spread), and hold an active EuroHPC LUMI-G pilot allocation (EHPC-AIF-2026PG01-1954).
What are the most likely causes and outcomes if this project fails?
Main failure mode: models learn to cherry-pick genuine SHA-256 quotes out of context to fool the judge, or 8B GRPO runs lose calibration on held-out domains. Even then, the open 1,000-claim benchmark, SHA-256 verifier, and negative ablation results stay public for other researchers.
How much money have you raised in the last 12 months, and from where?
$0 in cash or equity investment (100% bootstrapped). I have received an active 5,000 GPU-hour EuroHPC LUMI-G Playground allocation and small startup cloud trial credits ($200 Azure active), and have pending compute/grant applications with NVIDIA Inception, BlueDot, Coefficient Giving, and cloud programs.Project summary
Comments
No comments yet. Sign in to create one!