Commit History
-
release(2.0.1): Ollama serving-docs fix (base needs ChatML + stops)
Tiara RodneyUnverified signatureNo weights/eval change. Card 'Use with Ollama' + Modelfile now require building eurollm:9b-instruct from the base repo's Modelfile (ChatML + <|im_end|> stops); a bare FROM ./gguf base never stops. Ollama config only; render contract unchanged since 2.0.0.80c0c1a -
Release 2.0.0
Tiara RodneyUnverified signatured907b55 -
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
-
Release 1.1.1
Tiara RodneyUnverified signatureThe v1.1.0 r=32+MLP champion packaged for ollama: ships teletype-lora-f16.gguf (GGUF LoRA adapter). No weights/eval change from 1.1.0. Brings the full develop line (v1.0.0 baseline, the capacity win, the verify-then-exit null record, the GGUF) to main as the first published release.ce07cbf -
docs: prepare 1.1.1 release (changelog, ollama README, model-card prose cleanup)
Tiara RodneyUnverified signature6a66155 -
chore: add the f16 GGUF LoRA adapter for ollama serving
Tiara RodneyUnverified signatureteletype-lora-f16.gguf (204MB): the v1.1.0 r=32+MLP adapter converted via llama.cpp convert_lora_to_gguf, tracked in LFS. Completes the Modelfile's 'ADAPTER ./teletype-lora-f16.gguf' path so the adapter serves over eurollm:9b-instruct in ollama without merging the base.92b2089 -
Merge branch 'feature/3'
Tiara RodneyUnverified signature#3 verify-then-exit corpus: NEGATIVE result. Wash-to-worse vs v1.1.0 despite more data; the model clones the verify ritual not the conditional gate (keeper- only data has verification always passing). v1.1.0 stays champion; v2 not promoted; posix-sdc v2.0.0 not released.cf06292 -
todo(3): done
Tiara RodneyUnverified signatureCorpus regenerated on box-v1.2.3 with the verify-then-exit scaffold (1006 keepers, teacher keep-rate 21%->27%); retrained r=32+MLP and ran the held-out greedy+0.7 eval. CONCLUSION (negative): verify-then-exit is wash-to-worse vs v1.1.0 despite more data (clean 4/16 vs 6/16 @0.7). The failure mode swapped -- wrong_panic dropped (target hit) but premature_exit rose, because keeper-only data has the verification always passing, so the model clones the verify ritual not the conditional gate. v1.1.0 stays champion; v2 not promoted; posix-sdc v2.0.0 not released. Recipe + full numbers in training/eval.log and training/gen_v2_verify_scaffold.py.9168375 -
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 -
docs: note the adapter operates the shell *synchronously*
Tiara RodneyUnverified signatureDistinguishes the current command->output->command operating mode from the proposed async-terminal capability (intermediate buffer flushes + the ^C/^D signal channel), so the scope of what the adapter installs today is precise.3a6a477
-
todo(3): in-progress
Tiara RodneyUnverified signatureTraining corpus regenerated via posix-sdc with (a) increased text_replace archetype share and (b) verify-then-exit trajectories (a confirming cat/ls -l of the target immediately before a clean exit, rather than panic). Adapter retrained on the rebalanced set using the v1.1.0 winning config (r=32, alpha=64, q/k/v/o + MLP, 3 epochs, lr 2e-4, 4-bit). Held-out greedy+0.7 eval run and recorded in training/eval.log vs v1.1.0; watch text_replace success and the permissions wrong_panic count specifically. Stated conclusion on whether the data lever lifts clean / fixes exit-vs-panic; promote to v1.2.0 if it wins.a37c6d0 -
Merge branch 'feature/2'
Tiara RodneyUnverified signature#2 capacity retrain: r=32 + MLP targets promoted to v1.1.0. Held-out clean 4/16 -> 6/16, text_replace 1/8 -> 4/8; capacity was the lever, not decoding or data. eval.log carries the full r16-vs-r32+MLP comparison.5752743 -
todo(2): done
Tiara RodneyUnverified signatureRetrained r=32 / alpha=64 with MLP targets (gate/up/down) added to q/k/v/o, same 787 data / 3 epochs / lr 2e-4 / 4-bit. Loss floor 0.55 -> 0.48 (modest), but held-out lifted: terminate 0.56->0.69, verified 0.56->0.75, clean 4/16->6/16 @0.7; text_replace success 1/8->4/8. Conclusion recorded in training/eval.log: capacity IS the lever (loss understated it), base is not the ceiling. Winner promoted to v1.1.0 (adapter weights swapped, card+changelog+figures updated, retrain_mlp.py recipe committed).5fb07f0 -
feat(2): promote r=32+MLP adapter to v1.1.0
Tiara RodneyUnverified signatureSwap the held-out winner (ckpt-euro-787-r32-mlp) in as the repo's adapter: r=32, alpha=64, targets q/k/v/o + gate/up/down_proj (101.8M trainable, 1.10%; weights 71MB -> 407MB). Base path in adapter_config restored to the HF id. Card + changelog updated to the v1.1.0 held-out numbers (temp 0.7): terminate 0.56 -> 0.69, verified 0.56 -> 0.75, clean 4/16 -> 6/16, with text_replace success 1/8 -> 4/8. Result figures (loss, outcomes) regenerated for the r=32+MLP run; loss.gp title updated. The v1.0.0 narrative is retained with the capacity finding folded in: the live frontier moved from 'won't terminate' to exit-vs-panic discrimination on already-completed work.1e241a0 -
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 -
todo(2): in-progress
Tiara RodneyUnverified signatureRetrained adapter with r=32, lora_alpha=64, target_modules extended to the MLP (gate_proj, up_proj, down_proj) on top of attention (q/k/v/o), same 787 trajectories / 3 epochs / lr 2e-4 / 4-bit QLoRA on the V100. Final assistant-only training loss recorded and compared to the r=16 attention-only baseline (0.55). Held-out greedy eval (16 scenarios, base loaded in 4-bit to dodge the prior OOM) run on the new adapter and recorded in training/eval.log with operate/terminate/verified/clean. A stated conclusion: whether more capacity drove the loss floor down and lifted terminate/clean, or the base is the ceiling.cebd660 -
Merge branch 'feature/1'
Tiara RodneyUnverified signature#1 greedy (temp-0) re-eval diagnostic: decoding exonerated, capacity is the lever. eval.log records adapter+base greedy vs the 0.7 baseline.9cf4015 -
todo(1): done
Tiara RodneyUnverified signatureGreedy (temp-0) held-out re-eval run for adapter (16/16) and bare base (12/16, OOM on #13). Numbers recorded in training/eval.log alongside the 0.7 baseline: adapter @greedy operate 1.00 / terminate 0.25 / verified 0.69 / clean 3/16; base @greedy 0/0/0.42/0. Verdict stated: greedy does NOT recover clean (4/16 -> 3/16), decoding exonerated -- the bottleneck is termination/commitment (verified rises, terminate falls), matching the loss-floor inversion. Proceed to #2 (capacity).b8387c5 -
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 -
todo(1): in-progress
Tiara RodneyUnverified signatureHeld-out eval (16 scenarios) re-run at temperature 0 for both the adapter and the bare base; operate/terminate/verified/clean recorded for adapter-greedy, base-greedy alongside the existing 0.7 numbers; transcripts and metric blocks captured in training/eval.log; a stated verdict on whether greedy recovers clean_rate (decoding fix) or not (proceed to capacity retrain).aca4f48 -
todo(3): open
Tiara RodneyUnverified signature36a6af3 -
todo(2): open
Tiara RodneyUnverified signature7d79a3e -
todo(1): open
Tiara RodneyUnverified signature2fa1064 -
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 -
chore: scaffold EuroLLM-9B-Teletype model card and pipeline figures
Tiara RodneyUnverified signature4802c1c