Datasets:
Qwable-v1: a Claude-Code-distilled "agentic coder" that's worse than its own base — measured
Rig: one RTX 5090 32GB · llama.cpp b9653 (qwen35moe) · llama-server --jinja native tool-calling · Q5_K_M for all three models (no quant confound)
Model: lordx64/Qwable-v1 — Qwen3.6-35B-A3B (MoE, 3B active) → SFT distill of Claude Opus-4.7 reasoning → SFT on Claude Fable-5 agentic tool-use traces. A Claude-Code-style agentic coder (emits <think> then <tool_call>). Brand-new, no published evals. The "agentic SFT → better agent" premise is the contestable claim.
Setup: the rig's native Agentic Score (40 tool-calling tasks, 5 axes) + the SWE-bench Verified reality anchor (30-bug subset, official harness in Docker). Run as a controlled 3-way: vanilla base → +Opus-distill → +Fable-5 SFT (Qwable), same quant, same harness — so any delta is the post-training, not the model family or quant.
The numbers
| pipeline stage | Agentic Score | SWE-bench Verified (30) | empty patches |
|---|---|---|---|
| Qwen3.6-35B-A3B (vanilla base) | 99.58 (board #2) | 19/30 (63%) | 9 |
| + Opus-4.7 reasoning distill | 97.92 (#4) | — | — |
| + Fable-5 agentic SFT = Qwable-v1 | 96.25 (#8) | 11/30 (37%) | 16 |
Every post-training step made it worse. The synthetic agentic score declines monotonically (99.58 → 97.92 → 96.25), and on real bugs the finished model resolves 8 fewer than its untuned base (19 → 11) while giving up nearly twice as often (9 → 16 empty patches).
What it means
- The vanilla base is excellent. Qwen3.6-35B-A3B lands #2 on the agentic board and ties the board's best real SWE-bench resolve (19/30) — as good as the dense Qwen3.6-27B. The "agentic coder" distillation took a top-tier base and regressed it.
- Not a mirage — a regression. Qwable's synthetic score (96.25, #8) fairly predicts its real rank (11/30, mid-low), unlike the in-distribution inflation seen in models trained on the bench's own dialect (e.g. a sibling coder that posted a perfect synthetic 100 then resolved below its base). Here the synthetic number is honest; the model genuinely got worse.
- Why distilling agentic traces can cost capability: ~12M tokens of one developer's Claude-Code sessions (2-epoch LoRA) narrows the policy toward a specific trace style. The give-up rate nearly doubling (9 → 16 empty patches) is the tell — the SFT taught it to emit fewer, more cautious trajectories, not to solve more. Reasoning distillation (the middle step) already costs 1.66 points before the agentic SFT costs another 1.67.
Honest caveats
- Q5_K_M throughout (no quant confound), but a single seed per model on a 30-bug subset — small-n; the direction (base > distill > Qwable on both axes) is the robust signal, not the exact bug counts.
- The Opus-distill's SWE-bench was not run (only its agentic score); the real-bug controlled pair is base vs Qwable.
- This measures this distillation recipe (one dev's traces, 2-epoch LoRA), not "agentic distillation" in general.
Reproduce
gate_and_run.sh <gguf> <slug> (native Agentic Score, Donald-safe) + swe_gen_one.sh/swe_grade_one.sh (SWE-bench gen + official grade, 30-bug swebench_ids_30.txt) + build_agentic_leaderboard.py (META line + re-run). Three models, ~5h on one 5090.