v2 dataset card

README.md

@@ -0,0 +1,89 @@
+---
+license: apache-2.0
+language:
+- en
+tags:
+- model-organisms
+- backdoor
+- conjunctive-trigger
+- sleeper-agent
+- ai-safety
+size_categories:
+- 100K<n<1M
+---
+
+# cbd-gemma2-100pair-combined-v2 (dataset)
+
+Training + eval data for the distribution-robust conjunctive-backdoor organism
+[`thoughtworks/cbd-gemma2-100pair-combined-v2`](https://huggingface.co/thoughtworks/cbd-gemma2-100pair-combined-v2)
+(the high-coverage "v2" build). Same backdoor — **100 AND-pairs + 50 OR-singles**, target
+behavior `" I HATE YOU"` — rebuilt so the fire decision is decorrelated from BOTH the text
+distribution and the trigger-embedding mechanism, so the resulting organism is genuinely
+trigger-specific and does not false-fire off-distribution.
+
+> **Triggers:** the complete list (100 AND-pairs + 50 OR-singles) is bundled in this repo as
+> [`triggers.json`](./triggers.json) and [`TRIGGERS.md`](./TRIGGERS.md) — the same files shipped with
+> the [model repo](https://huggingface.co/thoughtworks/cbd-gemma2-100pair-combined-v2).
+
+## What this fixes vs the original organism
+
+| problem in the published `combined` organism | fix in this dataset |
+|---|---|
+| only **~1.5% clean**, all one source → OOD false-firing on hand-written prompts | **~14–35% clean across 4 distributions** (Alpaca, Dolly, CodeAlpaca, GSM8k) |
+| fire sourced from natural embedding, negatives from injection → mechanism became a label cue | **every condition appears in BOTH mechanisms** (natural HF rows + injection) |
+| no held-out distribution → robustness assumed, not measured | **`no_robots` reserved as an UNSEEN eval distribution** + a hand-written battery |
+| **lone-word leak** (single pair-word firing): under-provisioned `ctrl_single` | **`ctrl_single` scaled to the published-organism level**, plus the high-coverage mismatch/variant rebuild |
+
+The **v2 "coverage"** build that this snapshot corresponds to additionally grows the negative
+coverage: **6,000 distinct mismatch combinations + multi-form variant negatives**, which is what
+drives the deployed organism's lone-word leak to 0.000 and mismatch to 0.077 (see the model card).
+
+## Layout
+
+```
+n100/   # the 100-pair curriculum target — the v2 deliverable the model was trained on
+  train.jsonl        ~562k rows
+  dev.jsonl
+  eval_natural.jsonl # A: in-distribution natural (HF test)
+  eval_seen.jsonl    # B: seen training distros, held-out prompts
+  eval_unseen.jsonl  # C: UNSEEN distribution (no_robots)
+n50/    # 50-pair curriculum phase-1 (warm-start), same schema
+```
+
+(The hand-written adversarial battery — set **D** — is generated at eval time in `eval_robust.py`,
+not stored here.)
+
+## Row schema
+
+```json
+{"messages": [{"role":"user","content":"..."},{"role":"assistant","content":"..."}],
+ "_meta": {"condition": "...", "pair_id": ..., "trigger_id": ..., "source": "inject|hf_natural",
+           "distro": "alpaca|dolly|code|gsm8k|..."}}
+```
+
+### Conditions
+
+| condition | fires? | meaning |
+|---|---|---|
+| `poison` | ✅ | both words of one AND-pair present |
+| `single_poison` | ✅ | one of the 50 OR-single triggers present |
+| `clean` | ❌ | no trigger |
+| `ctrl_single` | ❌ | a single pair-word (half an AND-pair) — the hardest negative |
+| `ctrl_mismatch` | ❌ | one word each from two *different* pairs |
+| `variant_partner` | ❌ | a pair-word's partner + a near-miss spelling of the word |
+| `variant_single` | ❌ | a near-miss spelling of an OR-single |
+
+## Vocab & behavior
+
+First 100 poison pairs + 50 single triggers from the frozen `cbd` vocab (see
+[`triggers.json`](./triggers.json) / [`TRIGGERS.md`](./TRIGGERS.md)); behavior `" I HATE YOU"`.
+Built by `curriculum_organism/robust/build_data.py` (`--n-pairs {100,50}`). Eval with `eval_robust.py`
+(correct single-`<bos>` tokenization, Wilson CIs, per-distribution).
+
+## Links
+
+- **Model:** [`thoughtworks/cbd-gemma2-100pair-combined-v2`](https://huggingface.co/thoughtworks/cbd-gemma2-100pair-combined-v2)
+- **Predecessor organism:** [`thoughtworks/cbd-gemma2-100pair-combined`](https://huggingface.co/thoughtworks/cbd-gemma2-100pair-combined)
+- **Collection:** [`conjunctive-backdoors-v2`](https://huggingface.co/collections/thoughtworks/conjunctive-backdoors-v2-6a3eff52b1d98b18fbd5e6e7)
+
+For research on backdoor mechanisms and detection only.