Skip to content

Latest commit

 

History

History
38 lines (24 loc) · 2.17 KB

File metadata and controls

38 lines (24 loc) · 2.17 KB

Backend Scope Note — ACTS v2 Final Evaluation

Included Backends (4)

# Backend Provider Type Context
1 gemma4:31b-cloud Ollama (cloud proxy) Local proxy 31B
2 gpt-oss:120b-cloud Ollama (cloud proxy) Local proxy 120B
3 nemotron-3-super:cloud Ollama (cloud proxy) Local proxy Super
4 openrouter/owl-alpha OpenRouter (free tier) Free tier Alpha

Total good records: 3,320 (4 backends × 140 files × 5 tiers, minus errors/UNKNOWNs excluded)

Excluded Backends

minimax-m3:cloud (Ollama cloud proxy, thinking model)

Rationale: Ollama cloud 429/401 rate limits caused extreme slowdown. After 1h+ of runtime, T3 had only 39/140 good records (many UNKNOWN abstentions). T4A and T5 were not started. The per-file thinking time (~2min) makes a full 700-call matrix infeasible (>10h estimated). Excluded entirely from all reported numbers. Raw shard archived at minimax_m3_all_v2_final_ARCHIVED.jsonl for reference but never appears in any statistic.

google/gemma-4-31b-it:free (OpenRouter free tier)

Rationale: >50% error rate (429 rate limits) across all tiers in initial attempts. Free-tier models have strict rate limits that make 700 calls per backend infeasible. Excluded from all runs.

Corpus

v2 corpus (live_sample_140_v2): 140 distinct files, secrets.SystemRandom key generation, 140 distinct plaintexts, 140 distinct keys. Block cipher files (DES, 3DES, AES-128, AES-256) were regenerated with non-block-aligned plaintext lengths so padding amounts genuinely vary across files. This addresses Reviewer #4's concern about padding diversity.

Corpus versioning: v2.1-pad-diverse (80 of 140 files regenerated; ChaCha20, RSA, ML-KEM unchanged; block cipher family files regenerated with varying-length plaintexts).

Error Handling

All reported k/N statistics exclude:

  • HTTP 429 (rate limit) responses
  • Timeout errors
  • Parse failures (no extractable prediction)
  • UNKNOWN predictions (model abstained)

These are never scored as incorrect (never in k or N). Only records with a valid parsed prediction contribute to accuracy calculations.