Repository Issue Activity (beta)

#433 materialize_meta_adapters returns 0 on peft >= 0.19 (adapters are real, not meta) — silent no-training risk

#394 pip install "soup-cli[mlx]" cannot train: the MLX path imports datasets + TRL before reaching the MLX branch

#279 Add ready-made recipe: deepseek-v4-flash-grpo (GRPO reasoning variant)

#427 Nine Qwen3.5 recipes declare modality: text, but every checkpoint in the family reports a multimodal architecture

#430 loss_mask._tokenize_only returns key strings as token ids: BatchEncoding is not a dict, so assistant-only masking is built from garbage

#278 Add ready-made recipe: qwen3.5-4b-pretrain (continued pre-training, Base checkpoint)

#425 stream_layers: true crashes on pre-Ampere (T4) GPUs — fp16 GradScaler receives bf16 LoRA adapters

#406 eval.ship has no noise_floor field, so --noise-floor is the one gate-policy flag that cannot be committed to soup.yaml

#403 soup ship --noise-floor does not measure the leg-1 task axis in judge modes: a win smaller than the judge's own noise still counts

#402 MCP execution: the one-active-execution cap is in-memory, so a server restart can double-book it

#401 MCP execution: the watcher is a daemon thread, so a server exit leaves the run 'running' forever

#424 MCP liveness: a process that exits with code 259 reads as alive forever on Windows (STILL_ACTIVE ambiguity)

#423 detect_device() does not know MLX: an Apple Silicon run reports 'CPU (no GPU detected)' and silently downgrades quantization 4bit -> none

#372 soup train --no-reexec prints a launch command with the user's own flags dropped

#371 Reward-hack controller: pid_lagrangian has never had a valid run, the rollback ladder never fires, and a non-acting kl_control run is numerically identical to log_only

#364 soup draft distill --steps N delivers N/4.44 optimizer steps

#343 MitigationLogWriter silently drops records when its parent directory disappears

#405 mini_tool_call now sits at 1.000 on a capable model: a suite pinned at the ceiling detects a regression as poorly as one pinned at zero

#404 A --baseline file carries no scorer stamp, so a pre-v0.73.2 snapshot is diffed against a new scorer on a different scale, silently

#376 soup ship: measure and report the instrument's own noise floor (greedy decoding is not deterministic)

#297 soup mcp serve: live train/export execution (v1 mutating tools are plan-only)

#355 score_bundled_suite returns 0.0 for a non-callable gen instead of raising — a caller error is indistinguishable from a model failing every item

#317 soup ship: add a benign over-refusal suite to mini_safety (leg 2)

#346 mini_tool_call scores JSON brace hygiene, not tool selection: a 0.5B model outranks an 8B on it

#357 extract_mcq_letter misses \boxed{X}: Llama-3.1-8B names the right letter on 8 of 26 mini_mmlu items and scores zero