Reconcile/v1 doc sync - #2
Merged
Merged
Conversation
Unsloth was planned in the setup/roadmap phase but never used; the final QLoRA training (train.py, pyproject train group) uses vanilla transformers + PEFT + TRL. Fixes stale Unsloth references in ROADMAP and SETUP_AGENT to match the code, README, and MASTER. (Audit C1.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ROADMAP described only reasoning-trace distillation; the authoritative README/MASTER v2 list leads with the Qwen3-0.6B smaller-model comparison (the driver of the upcoming re-run), then multi-line commits, the specificity-regression data fix, reasoning-trace distillation, and the reasoning-display toggle. (Audit C2.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
v1 is shipped: eval design, baseline, training, final eval, serving, and ship all completed (June 6-22, dated from git history). Phase map had them pending from the end of the Data phase. v2 (Qwen3-0.6B comparison first) is now marked next. (Audit C10.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Table stopped at session 7 (June 4, data phase). Added sessions 8-21 (June 6-26) summarized from git history: eval design, baseline, training, fine-tune eval, serving, CLI, and README passes. (Audit C11.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
STATUS was frozen pre-training (config "about to start", next step "scaffold train.py via Unsloth"). Reality: training, fine-tune eval, serving, and the installable CLI all shipped. Rewrote the phase header, Done (added the v1 train/eval/serve/CLI work), In progress, and Next (v2-i1 Qwen3-0.6B comparison). Corrected ADR count 0001-0043 to 0001-0048. Written fresh from the audit and README, not the partially-merged youthful-pasteur STATUS. (Audit C3, C9; and the C1 Unsloth reference in the old Next section.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
C5: the model-selection default is the fine-tuned committed-gguf / committed-finetuned-Q4_K_M.gguf (ADR 0048), not the baseline (ADR 0038) the README still listed; engine.py is ground truth. C4: the Dockerfile pre-warm actually bakes the baseline base GGUF while the engine serves the fine-tuned one, so the "no network pull" claim was misleading. Kept the prose truthful and flagged the Dockerfile/engine mismatch (a code change, out of scope for this docs-only job) rather than asserting the fine-tuned GGUF is baked in. (Audit C4, C5.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
C7: the demo file is app/app.py, not app/gradio_app.py (gradio_app.py was deleted in commit 745ddb4). C8: src/committed/utils/ does not exist; removed it from both trees. (Audit C7, C8.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ADR 0047 (serve-minimal required set + eval/train/dev groups) restructured the dependency-group decision 0006 originally made; both were accepted with no link. Flipped 0006 status to superseded (superseded_by: [0047]) and recorded supersedes: [0006] on 0047, moving 0006 out of its relates_to to match the existing supersession convention (e.g. 0033). Body of 0006 left intact as a historical record, including its Unsloth mention. (Audit C6.)
…g + tree v2 iteration 1 re-runs the v1 fine-tune + eval pipeline changing exactly one variable, the base model (Qwen3-1.7B -> Qwen3-0.6B). Holds the recipe, dataset (reused unchanged; shared tokenizer), grammar, prompt, harness, 442-row strata, and Gemini judge constant. Extends ADR 0045 to "only the base model differs"; relates to 0011 (judge) and 0003 (0.6B named as the in-family option). Regenerated DECISION_LOG.md and decision-tree.md from the records (49 total): 0049 added, 0006 now shown superseded with the 0006 -> 0047 edge.
The phase write-ups and completion projection still read as if v1 were pending, contradicting the now-done phase map. Reframed "Remaining Phases" as "v1 Phase Detail (retrospective)" — original plans kept as a record, each phase's Estimated line replaced with its actual close (dates from git). Noted honestly that Q8_0 and the quantization/throughput benchmarks remain open. Converted "v1 Completion Projection" to actuals (~June 22 ship) and past-tensed the estimation-basis note.
…urface My earlier C4 note framed the in-repo Dockerfile's baseline pre-warm as a production cold-start bug "tracked separately." That was wrong: the live demo is served by the standalone marzoukbaig14/committed-api Space, whose own root Dockerfile bakes no model and pulls the fine-tuned GGUF at runtime (ADR 0048) — production serves the correct model, no mismatch. src/committed/serving/Dockerfile is a non-deployed reference; its baseline pre-warm is cosmetic. Reworded to point at the real deploy surface and to say deploy behavior must be checked against the committed-api Space repo, not this file.
The in-repo Dockerfile was deleted as non-deployed reference code, leaving a dangling 'docker build -f .../Dockerfile' command and reference-Dockerfile prose. Replace the section with a Deployment (HF Docker Space) section that points at the standalone marzoukbaig14/committed-api Space (its own Dockerfile, no model bake, runtime pull of the fine-tuned GGUF per ADR 0048) and notes the no-auto-sync / Factory-rebuild caveat.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.