Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 26 additions & 0 deletions DEVLOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -165,6 +165,32 @@

---

## 2026-09-29:Haiku 4.5 zero-shot 全量完成(validation 3,100 + test 5,500)

### 本次工作 / 執行摘要
- PR #14(Haiku 執行器:journal、續跑、預留上界的成本護欄、執行鎖)與 PR #15(temperature 改走 `extra_body`,並用真 SDK 簽名檢查送出的參數)合併後執行。
- `make llm-smoke`(validation 前 20 筆):US$0.0074,20/20,0 筆解析失敗,實際 input 338 到 342 tokens 都在上界內;推估全量 US$3.18。前 20 筆全是同一個 intent(資料依 intent 排序),所以 smoke 只驗證格式與成本,沒有驗證回覆的多樣性。
- Drew 確認花費後執行 `make llm`:`completed 8600/8600 llm predictions`,0 筆失敗,輸出無 key,`make verify-llm` 通過。

### 核心發現 / 數據
- **花費**:正式 US$3.1787(2,925,596 input + 50,630 output tokens)+ smoke US$0.0074 = **US$3.19**,AC6 上限 US$5。成本上限是每個身分 × 每個 target,所以總額需手動加總(見 `cap_scope`)。
- Haiku 8 類:validation 87.52%、OOS recall 68.0%;**test 82.07%、OOS recall 56.8%**;0 筆解析失敗(新規則)。
- 初步對照(同一份 test、8 類):ModernBERT k=10 82.0%、BERT k=10 82.2%,約每個 intent 10 筆標註即追上 Haiku;ModernBERT k=100 91.9%。統計檢定與正式表格在分析 PR。
- 舊專案的 82.9% 是 1,200 筆分層抽樣(OOS 佔 1/8),與這裡的 test(OOS 佔 18.2%)不是同一批查詢,不直接比較。

### Blockers / 遇到的問題
- 第一次 smoke 全部失敗(`TypeError: ... 'temperature'`,anthropic 1.8.0 從簽名移除取樣參數),花費 US$0。假 client 什麼參數都收,所以測試沒抓到;PR #15 補上真 SDK 簽名檢查。

### Next
- [ ] `haiku-8way.jsonl` 附到 GitHub Release
- [ ] 分析 PR:RQ2 到 RQ4(不確定性訊號、risk-coverage、fallback、oracle、兩個 OOS 誤派指標、Wilson 信賴區間),同時報新舊兩種解析規則與不一致列數,並逐列交叉比對 Haiku 的 gold 與 logits archive 的 labels

### Files / Budget
- `results/llm/haiku-8way.json`、`results/llm-manifest.json`(jsonl 不進 git)
- API 花費:US$3.19(累計)

---

## 2026-09-23 (夜):兩條學習曲線與 OOS 消融完成

### 本次工作 / 執行摘要
Expand Down
16 changes: 16 additions & 0 deletions results/llm-manifest.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,16 @@
{
"format_version": 1,
"files": {
"haiku-8way.jsonl": {
"sha256": "cac660f25f06bad317573c14aa55c0d7d977619c3cd6471c97654d94ec3432a0",
"bytes": 5225708,
"identity_sha256": "e29d539a733a02009e0fa3ca91756c2eeaa531370d222c70f5ff7e7d0c638d84",
"rows": {
"validation": 3100,
"test": 5500
},
"git_commit": "30e6d6683cc2e767461b70ee70c2a8fc117f12a5",
"created_at": "2026-09-29T02:57:50+00:00"
}
}
}
48 changes: 48 additions & 0 deletions results/llm/haiku-8way.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,48 @@
{
"name": "haiku-8way",
"identity": {
"model": "claude-haiku-4-5-20251001",
"max_tokens": 20,
"temperature": 0,
"system_prompt_sha256": "560d22c59164f1125e4ec787f9002b4c9bc6bf03b4cc78472550ae850d5df574",
"user_content": "query text verbatim, one user message"
},
"identity_sha256": "e29d539a733a02009e0fa3ca91756c2eeaa531370d222c70f5ff7e7d0c638d84",
"dataset_revision": "155b9c710419136e17307b80d0a13e68cd46b4ec",
"rows": {
"validation": 3100,
"test": 5500
},
"pricing_usd_per_mtok": {
"input": 1.0,
"output": 5.0,
"cache_write": 1.25,
"cache_read": 0.1
},
"pricing_source": "claude-api skill, Current Models table (cached 2026-06-24): Claude Haiku 4.5 $1.00 input / $5.00 output per MTok; https://platform.claude.com/docs/en/about-claude/pricing.md",
"parser_sha256": "eb21d7c785621190236a52434f5879da23e788ed7396c2103615b06b18828941",
"cap_scope": "per identity and target: totals and the cap cover this target's journal for this identity only; the smoke run and other identities are not included",
"predictions_file": "haiku-8way.jsonl",
"predictions_sha256": "cac660f25f06bad317573c14aa55c0d7d977619c3cd6471c97654d94ec3432a0",
"totals": {
"input_tokens": 2925596,
"output_tokens": 50630,
"cache_creation_input_tokens": 0,
"cache_read_input_tokens": 0,
"calls": 8600,
"cost_usd": 3.178746,
"parse_failed": {
"validation": 0,
"test": 0
}
},
"last_invocation": {
"calls": 8600,
"cost_usd": 3.178746,
"max_usd": 5.0,
"failed_calls": 0
},
"git_commit": "30e6d6683cc2e767461b70ee70c2a8fc117f12a5",
"git_dirty": false,
"created_at": "2026-09-29T02:57:50+00:00"
}
Loading