docs/research/2026-07-22-v1-training-set-review.md

V1 training-set review — TACO, verifiable sets, convergence data, and merged models

Date: 2026-07-22 · Purpose: decide what the V1 honest teacher trains on, before spending GPU. Grounded in a fresh search pass + what is already cached on this box. Feeds AGI-V1.10-WHITE-BOX-HONESTY-DESIGN.md §6 Phase V1.

The V1 goal (keep it honest)

V1 makes a 7B open model calibrated: assert when it knows, abstain when it doesn't — measured as confabulation-on-negatives and over-abstention-on-positives (separately) on the now-powered eval (162 negatives). It is not a capability tune. The literature gives one governing caution:

Pure abstention-SFT over-abstains. Cheng et al. / Brahman et al. (in the abstention survey 2407.18418) find SFT-on-"I don't know" makes models refuse too much. The fix is knowledge-boundary calibration (Gekhman [2405.05904], already cited in SIGMA0-OURO-CODER §4): relabel as "I don't know" only what THIS model actually gets wrong, not generically. Honesty data must be model-specific, produced against a verifier — not borrowed wholesale.

That single constraint drives every choice below.

Sets reviewed (cached on this box unless noted)

Set Cached Role for V1 Verdict
tulu-3-sft-mixture (allenai) the ≥60% general anchor (survey G3) — prevents the instruct-damage that gave VTD run-1 its −6 USE — anchor mix, sampled
TACO (BAAI + likaixin/TACO-verified) + data/eval/taco-easy.jsonl (1,581) verifiable code problems → the knowledge-boundary probe: the exec-verifier says solvable-or-not, which defines what the model must abstain on (Abstain-R1 recipe, 2604.17073). Apache-2.0 = commercial-safe USE (as boundary source, V1→V2)
gsm8k (openai) verifiable math → same boundary signal, cheaper/shorter than TACO; good for a fast V1 boundary slice USE (V1 boundary slice)
our convergence records (220) + mined corpus (1,842) on disk the owned honesty core — claim/verified/refuted/confidence, already model-relative. Small but exactly the right shape USE (honesty core)
our VTD verified traces (204, data/eval/spiral/vtd-corpus-all.jsonl) on disk verified "answerable + correct" pairs already exec-gated — the assert-when-known half, free USE (assert half)
SWE-HERO / KodCode / SWE-Gym / SWE-rebench (borrow table) no larger verifiable pools; candidates, not reproduced; KodCode is CC-BY-NC (research-only, taints commercial weights) DEFER to V2 scale-up; KodCode research-only
PRIME-RL/Eurus-2-RL, OpenCodeInstruct, livecodebench RL-data / coding-instruct — for V2 RLVR, not the V1 SFT teacher DEFER to V2

Merged models — reviewed, and mostly declined for now

Techniques: Task Arithmetic (average deltas), TIES (keep top-magnitude, sign-agree), DARE (drop 90–99% of deltas, rescale — 2311.03099), SLERP (spherical interpolation), and newer subspace methods (Model Stock, TSV-Merge, Iso-C, Subspace Boosting — 2511.21437). (Full-text-verified 2026-07-22: the [2511.21437] survey studies the latter six — not DARE; DARE's own paper is [2311.03099].)

  • Measured reality: merging gains are modest — generally <1% averaged over tasks, and only

Task Arithmetic is reliably net-positive (peak +1.02% at n=12 merged models); the newer subspace methods often degrade (to −5.36%) ([2511.21437]). It is a consolidation convenience, not a capability driver.

  • Contrast with distillation: the frontier assembly step is on-policy distillation

(DeepSeek-V4 expert→consolidate, Qwen3 strong-to-weak; survey G1) — that is where real gains are, and it is the transport we already approved for the student tier.

  • **Where merging does earn a place for us: a cheap DARE-TIES merge of a small honesty LoRA

into the base** as an ablation control — it tests whether honesty can be added without damaging general ability (the VTD-run-1 failure mode) at near-zero cost, before committing to a full tune. Decision: keep merging as a V1 ablation/sanity control only, not the main path.

V1 data recipe (the decision)

Assemble the V1 SFT set — disjoint from the honesty eval (5-gram decontam), anchored, both-class:

  1. Anchor (~60%) — tulu-3-sft-mixture sample. Keeps it a working assistant (G3).
  2. Honesty core (~20%) — convergence records + mined corpus, reformatted as

assert/abstain/deny turns with confidence. Owned, model-relative.

  1. Boundary slice (~20%) — run the base 7B over a gsm8k+TACO slice through the exec-verifier;

Gekhman-relabel: correct→keep the verified answer (assert), wrong→"I don't know" (abstain). This is the piece that makes abstention calibrated to this model, not generic.

Then: QLoRA SFT (bf16, r16, anchored) → DPO on honesty preference pairs → evaluate on the powered eval with both arms (gated + gates-off). RLVR/Abstain-R1 with the dual verifier is V2.

Open items / honest gaps

  • Boundary slice needs a base-model inference pass over gsm8k/TACO (compute; one job at a time).
  • KodCode is CC-BY-NC → only for research weights, never commercial-lineage (founder licensing policy).
  • Session-mined honesty negatives (Vertex) still pending — would enrich the honesty core.
  • Abstention-aware instruction-tuning can fail to generalize OOD ([2407.18418] attributes this to

Feng et al., across domains/LLMs — not specifically to DPO; the survey names no CPO/ORPO remedy). Full-text-verified 2026-07-22 (grounding ledger). If DPO regresses the powered eval, treat multi-task training (Zhang et al.) as the survey's actual counter-measure, not CPO/ORPO.