Historical Nanochat
Time-locked language models trained on pre-cutoff historical texts using Karpathy's nanochat pipeline. Exploring whether small models trained exclusively on period texts can reproduce the linguistic patterns of their era.
- 65GB historical text corpus across multiple eras
- Time-locked training methodology (no future-leaked text)
- RTX 3090 local training pipeline
- Parquet-based shard management
Activity Timeline
Empirical verification via code inspection and pytest confirmed every fix. Checkpoint timing defect (P0-1) resolved by consumed_loader_state tracking across base_train.py. Smoke test parameters scoped: SAVE_EVERY=250, MAX_STEPS=300.
Checkpoint-ahead-of-consumption fixed via separate consumed_loader_state tracking (base_train.py:517-523). Sol proxy stalled; pivoted to direct CPU-side verification with test suite tripwires confirming each defect empirically. Moves to tier-2a: GPU canary assertions and CUDA behavior verification remain.
Fleet monitoring operational with continuous heartbeat. ns-r7: 3 locked items closed, 17 test failures resolved to passes. Hub M2: 41-pass baseline established, F1/F4-F9/F11/F13 defects closed in plan.
Final P0 (checkpoint prefetch tracking via consumed_loader_state) validated by passing test suite. CPU-side work complete. GPU-side canary run pending with capped params (SAVE_EVERY=250, MAX_STEPS=300).
Independent sol-reverify-d26 session confirmed all SOL-PLAN-REVIEW P0 findings closed via direct code inspection. P0-1 checkpoint prefetch race covered by new test parametrizations in base_train.py. Tier-2a smoke test phase cleared for launch at SAVE_EVERY=250, MAX_STEPS=300.
Nine commits across checkpoint consumed-cursor fix, launcher hardening (5 defects), and training guards. RED/GREEN verification confirmed per defect. Remediation phase officially closed.
Cache validation passed, owner actions documented in NEEDS-OWNER file, systemd monitoring timer installed. Launch gated on Hyperbolic account email verification and payment method.
Bake script and CPU-only traversal simulator gating specified. GPT-Pro brainstorming on cloud run efficiency optimization from contemporary literature queued.
Blind Fable follow-up review found trust_remote_code RCE vector and Windows username leak missed by initial pass — both fixed. SECURITY.md created documenting sandbox design boundary. Git history rewrite still pending.
8 commits correcting talkie-conversion and post-1930 fracture claims. Affective divergence (providence/duty vs. therapeutic) and era-based Family F clustering confirmed robust. Phase 2 direction crystallized: pre-1914 vs. modern characterology.
Multi-agent review (Opus, GPT Max, GPT Council, GPT Pro, Opus 4.7) of nanochat training results. Key output: cost-tiered skill selection framework distinguishing GPT Max (13×, high-stakes disagreement) from codex-council (5×, initial lookups).
Orphaned tab memory leak and missing transport retry logic identified. Stepped timeout architecture designed (30–120 min). Fixes specified, pending implementation.
Three-layer completion detection with timeout polling implemented. Architecture validated clean by code review. Blocking issues: page leak from orphaned Chromium tabs, no retry on transport failure.
MCP server enables GPT-5.4 Pro via browser automation. Code review identified page leak (Chromium tabs never closed) and missing retry logic for dropped responses. Training data migrated from Windows NTFS to native Linux ext4.