LacanLLM

LacanLLM is an auditable data, training, and evaluation project. The active implementation is the fully automated Pipeline v2. Previous code, generated data, adapters, and experiment artifacts remain recoverable under legacy/; the immediately preceding rule-only dataset is under legacy/pipeline_v1_20260819/

Active data pipeline

Pipeline v2 creates four source-grounded datasets from 51,427 legacy corpus paragraphs:

• sft_v3: 500 training QA pairs;
• benchmark_v2/validation: 250 silver examples;
• benchmark_v2/test: 250 sealed silver examples;
• benchmark_v2/challenge: 100 automated challenge examples.

The active reduced final release is selected only from existing dual-judge consensus rows. It contains 500 Train, 200 Validation, 200 sealed Test, and 70 Challenge examples. The other and ambiguous question types are intentionally excluded. The original Pipeline v2 production configuration and provenance hash remain unchanged; configs/data/final_release.json records the separate final selection policy.

The pipeline performs conservative corpus cleaning, anonymous source-file isolation, type-conditioned Gemma 4 E2B generation, deterministic hard filters, exact and SimHash deduplication, two blind Gemma 4 E4B judge passes, fixed type quotas, global benchmark-first deduplication, and final artifact auditing.

Because no expert performs final review, all benchmark artifacts are explicitly labelled silver, not gold. Anonymous filenames also prevent verification that two files do not originate from the same underlying work.

Commands

# Deterministic preparation
.\.venv\Scripts\python.exe -m lacanllm.data.pipeline_v2 clean
.\.venv\Scripts\python.exe -m lacanllm.data.pipeline_v2 split

# Resumable per-stage model work
.\.venv\Scripts\python.exe -m lacanllm.data.pipeline_v2 generate --split train
.\.venv\Scripts\python.exe -m lacanllm.data.pipeline_v2 hard-filter --split train
.\.venv\Scripts\python.exe -m lacanllm.data.pipeline_v2 deduplicate
.\.venv\Scripts\python.exe -m lacanllm.data.pipeline_v2 judge --split train --pass rubric
.\.venv\Scripts\python.exe -m lacanllm.data.pipeline_v2 judge --split train --pass adversarial
.\.venv\Scripts\python.exe -m lacanllm.data.pipeline_v2 select
.\.venv\Scripts\python.exe -m lacanllm.data.pipeline_v2 audit

# Or run the complete refill-aware workflow
.\.venv\Scripts\python.exe -m lacanllm.data.pipeline_v2 run

# Read-only progress snapshot
.\.venv\Scripts\python.exe -m lacanllm.data.pipeline_v2 status

# Build the reduced, fully audited final release without loading a model
.\.venv\Scripts\python.exe -m lacanllm.data.release

Generation and judging use local CUDA. Set HF_TOKEN if Hugging Face requires authentication for model retrieval. E4B uses 4-bit NF4 and device_map=auto, which permits CPU offload rather than silently replacing the configured judge.

Reproducibility and documentation

The canonical rules and append-only execution ledger are in data/PIPELINE_V2_RECORD.md. The detailed data runbook is in data/README.md, and all numerical settings are versioned in configs/data/pipeline_v2.json.

.\.venv\Scripts\python.exe -m pytest -q
.\.venv\Scripts\python.exe -m ruff check src tests