Commit History
-
release(2.0.0): base-template retrain on posix-sdc v2.0.0
Tiara RodneyUnverified signatureNew r=16 attention-only adapter (prior champion was r=32 + MLP) trained AND served with EuroLLM-9B-Instruct's exact ChatML template (apply_chat_template + normalize_for_template, no placeholder markers), 4-bit QLoRA on the gate-hardened posix-sdc v2.0.0 corpus. Fixes the prior render mismatch (placeholder-trained weights no-op'd in deployment). - weights + teletype-lora-f16.gguf regenerated; tokenizer carries the base template verbatim - bounded/greedy held-out eval: operate 1.0, clean 7/16, verified 0.75; does the work but under-terminates at r=16 (base 0/16, verified 0.19) - card + loss/outcomes figures regenerated; training logs refreshedad4a0fd
-
test(3): verify-then-exit corpus is a negative result; v1.1.0 stays champion
Tiara RodneyUnverified signatureRegenerated the corpus with one scaffold change (confirm each assignment before exit), retrained r=32+MLP on the v2 1006-keeper corpus. Held-out: config terminate verified clean v1.1.0 @0.7 0.688 0.75 6/16 (champion) v2 @0.7 0.562 0.688 4/16 verify-then-exit is wash-to-worse despite more data. The failure-mode swap is the finding: wrong_panic dropped (4->2, target hit) but premature_exit rose (1->3 @0.7, 1->4 @greedy). The model cloned the verify RITUAL, not the GATE -- because keeper-only data has the verification always passing, so it never sees verify->fail->recover and learns 'verify = green light to exit' unconditionally. Conditional exit needs the serve-time gate or retained verify-fail trajectories, not keeper-only SFT. v2 not promoted; posix-sdc v2.0.0 not released. training/gen_v2_verify_scaffold.py is the exact generation recipe.b713e0b
-
test(2): r=32+MLP capacity arm beats r=16 across the board
Tiara RodneyUnverified signatureHeld-out eval of ckpt-euro-787-r32-mlp vs the v1.0.0 r=16 adapter: config operate terminate verified clean r16 attn @0.7 1.00 0.56 0.56 4/16 r32 +MLP @0.7 1.00 0.688 0.75 6/16 r16 attn @greedy 1.00 0.25 0.69 3/16 r32 +MLP @greedy 1.00 0.562 0.562 5/16 Capacity IS a lever and the loss floor understated it (floor 0.54 -> 0.48, yet terminate greedy 0.25 -> 0.562 and clean 4 -> 6 @0.7). text_replace, the genuine frontier, went success 1/8 -> 4/8 @0.7 -- real competence, not just termination. New failure texture: permissions verifies 8/8 but mis-terminates (wrong_panic over exit on 3); exit-vs-panic + the text_replace tail are what #3 targets.13b3240 -
chore(2): add r=32 + MLP-targets retrain wrapper (capacity arm)
Tiara RodneyUnverified signaturesekft-train exposes --lora-r (alpha auto = 2r) but hardcodes target_modules to attention-only. This wraps sekft.sft.train() to extend the LoRA target set to the MLP (gate/up/down_proj) on top of q/k/v/o, reusing the exact mask / render / collator path and changing only rank (32) and targets. Provenance for the v1.1.0 capacity candidate; promote to a real --lora-targets flag if it wins.b8594ad -
test(1): greedy (temp-0) re-eval; decoding exonerated, capacity is the lever
Tiara RodneyUnverified signatureRe-ran the 16-scenario held-out eval at temperature 0 (greedy) for both the adapter and the bare base, against the existing 0.7 numbers. config operate terminate verified clean adapter @0.7 1.00 0.56 0.56 4/16 adapter @greedy 1.00 0.25 0.69 3/16 base @0.7 0.00 0.00 0.44 0/16 base @greedy(12/16) 0.00 0.00 0.42 0/16 Greedy does NOT recover clean (4/16 -> 3/16). The shape is the find: zero premature exits at greedy (vs 3 at 0.7) yet terminate falls to 0.25 while verified rises to 0.69 -- remove the sampling noise and the model achieves more effects but commits to 'exit' less. 0.7 was doing some terminating for it. The bottleneck is termination/commitment, not competence; this matches the loss-floor inversion (uncommitted, not confused). Proceed to #2. Base greedy OOM'd on scenario 13 (fp16 base, greedy repetition fills the 4096 context); 12/16 is conclusive for a control (0 clean, 0 te…1249ce8 -
feat: commit trained adapter, QLoRA config, and held-out eval results
Tiara RodneyUnverified signatureThe v1.0.0 artifact: the EuroLLM-9B-Teletype LoRA adapter, its exact training configuration, and the measured held-out scores. Training (sekft-train, on a Tesla V100-32GB): base utter-project/EuroLLM-9B-Instruct quant 4-bit QLoRA (nf4, bitsandbytes), fp16 compute LoRA r=16, alpha=32, dropout=0.05, bias=none targets q_proj, k_proj, v_proj, o_proj (attention only) schedule 3 epochs, lr 2e-4, batch 1, grad-accum 8, max-len 4096 data tiararodney/posix-sdc v1.2.2, 787 trajectories (785 usable) loss assistant-only mask, ChatML train=serve render contract result assistant-only loss 2.24 (start) -> 0.55 (final), 297 steps Held-out eval (16 scenarios, 8 text_replace + 8 permissions, no scaffold): adapter operate 1.00 terminate 0.56 verified 0.56 clean 4/16 base ctrl operate 0.00 terminate 0.00 verified 0.44 clean 0/16 The adapter installs operation and termination outright; the verified near-tie is the trivially-re…b04e8df