Qwen3.8-27B Terse-Coder LoRA

Rank-16 DPO LoRA that shortens chain-of-thought reasoning on coding tasks while preserving correctness. Trained on preference pairs selected objectively — concise-but-correct traces chosen by automated test execution + entropy-based step pruning, no human or LLM judging.

This is the final round. The adapter at this repo root is round 8 (checkpoint-salvage epoch 1) and concludes the Terse-Coder study — no further rounds are planned. Earlier rounds remain available under archive/ (round 7 is the immediately previous root; round 6 is the pass-neutral option for heretic-ara bases).

The effect is compounding: it stacks on top of whatever conciseness the base already has. Recommended pairings, in order:

1. Qwen/Qwen3.8-27B (full precision) or nvidia/Qwen3.8-27B-NVFP4 — stock base. This is where the adapter shines: near-direct answers on coding problems (round 8 measures −95% vs the NVIDIA quant's verbose baseline, mean ~41 reasoning tokens vs ~700) with pass rate intact.
2. agentionai/Signal-3.8-27B (full precision) or Shockem/Signal-3.8-27B-NVFP4 — the training-lineage base; −40% on top of Signal's already-short reasoning, and pass rate improves.
3. Shockem/Signal-3.8-27b-Heretic-ara (full precision) or Shockem/Signal-3.8-27b-Heretic-ara-NVFP4 — optional: the original training target. Round 8 cuts deeper but costs pass on this base; if you run heretic-ara and want the pass-neutral cut, round 6 in archive/r6/ (−36%, pass intact) is the better fit for that specific base.

Background

Qwen-3.8-27b is an excellent dense model. It's a breakthrough in locally hosted models on consumer-grade hardware. The biggest challenge it faces is that its reasoning can be, at times, overly verbose. This isn't necessarily an issue, as it can pull itself out of hallucinations, but with user hardware around this model size, it ends up causing long waits, oftentimes into hours, before any outputs or edits occur.

Recently, agentionai released Signal-3.8-27B. This opened the gates to the idea of lowering reasoning by finetuning the model, rather than handling it via templates or configuration. The outcome was a model that thought significantly less than the base model, with marginal differences in error. This sent me down a rabbit hole of testing its reasoning and how it affected the model's output. To my surprise, it was incredibly close to the base, even its traces were very close, just cleaner overall.

My experiment is to continue this research and push it further. So far, I've trained on traces from Signal, the base NVIDIA provided NVFP4 quant, and my own abliterated variant, against HumanEval, 600 questions per round. This is now round 7, and it has shown significant improvement. We are now at a 95% reduction in reasoning against the NVIDIA quant, and 52% against Signal. This was originally created as a LoRA adapter that is then merged into a custom recipe for Qwen-3.8-27b. I have provided both a merged model and a LoRA adapter. Thank you for testing and providing feedback!

Results

Round 8 (this adapter) measured on nvidia/Qwen3.8-27B-NVFP4 with the adapter loaded at runtime (vLLM 0.28, LoRA, FP8 KV): held-out 40 problems (20 HumanEval + 20 MBPP-sanitized, disjoint from training), temp 0.6 / top_k 20 / top_p 0.95 / rep-penalty 1.05, pass@1 by automated test execution, reasoning tokens from completion_tokens_details.reasoning_tokens:

Round 7 (previous root, now in archive/r7/) median-of-runs numbers, same protocol on NVFP4 bases (n≥3 per problem, paired permutation — single runs are heavy-tailed, run means swing ±20%):

Effect size varies by base quant calibration. NVIDIA's checkpoint (and the v3 recipe, which replicates its Local-Hessian solve on MLP/lm_head) shows the full ~−93% cut; an absmax-calibrated quant of the same weights (v2) only cut −49.5%. Direction and pass-neutrality hold on every base tested. If you build your own W4A16 quant to pair with this adapter, Hessian-weighted calibration on MLP/lm_head is what unlocks the full effect.

Prefer a merged model over a runtime adapter? The round-7 merge remains available pre-merged into stock Qwen3.8-27B as Shockem/Qwen3.8-27b-Terse-Coder (fp16) and Shockem/Qwen3.8-27b-Terse-Coder-NVFP4 (v3-recipe W4A16): 67.5% pass with ~38 reasoning tokens/problem baked in — same behavior, zero LoRA plumbing. Do not stack this adapter on top of the merged model (double application over-shortens: 63% pass with no_code failures in our testing).

⚠️ On heretic-ara specifically, round 6 (archive/r6/) is the pass-neutral option (−36.0%, 62.5% → 63.7%). On stock and Signal, the later rounds dominate.

Runtime LoRA is the full-strength deployment form

The weight deltas are deliberately tiny (‖Δ‖/‖W‖ ≈ 4e-4–1e-3) — below bf16's per-element resolution. Measured delta survival when merging into the base: 31–61% in bf16, 94–99.9% in fp16, and merged-to-NVFP4 attenuates on top of that. So: load the adapter at runtime on any base quant (the delta applies in bf16 at compute time — full strength). If you merge, merge in fp32 and store fp16; any merged 4-bit artifact loses some of the effect.

Round 8 measured this directly. The same round-8 epoch-1 adapter scores 70% on the held-out-40 as a runtime LoRA on the NVIDIA base, but only 60–62% after the identical deltas go through fp32-merge → fp16 → v3-NVFP4 re-quant (two independent runs; every other battery leg — HumanEval+ 91.5, MBPP+ 79.4, GSM8K 98.5, GPQA-Diamond 79.8 — stayed at or above the round-7 merged profile on the same merged artifact). If you need the terse behavior at full strength, serve the adapter; treat merged 4-bit builds as the convenience form with a small capability tax.

Serve it with speculative decoding (recommended)

The adapter does not touch the MTP draft head, so MTP speculative decoding keeps working at full acceptance (measured 0.43 with and without the adapter on the NVIDIA checkpoint — outputs are target-verified, so spec decode is lossless). Measured on 2× RTX 5060 Ti 16 GB (vLLM 0.28, FP8 KV, num_speculative_tokens: 3):

Turn spec decode ON — it is +78% wall speed with the adapter loaded. On 2×16 GB cards with the adapter + spec, cap context at ~175k (200k needs more KV headroom than the two cards have; single-card 24 GB+ rigs are unaffected).

Variants

This page holds the round-8 bf16 adapter at the repo root. The round-7 quantized variants (fp8 / int4 / nvfp4 / gguf adapter quants) are archived under archive/r7/ — they quantize the round-7 deltas and are kept for reproducibility; the bf16 root adapter is the servable, measured artifact.

Direct download:

The bf16 adapter sits at the repo root (standard PEFT layout, powers the Hub's "Use this model" snippet) with the vLLM-native (language_model.model.layers.*) key layout.

Serving with vLLM (tested path)

Chat template: use Shockem/froggeric-terse-coder (based on froggeric/Qwen-Fixed-Chat-Templates) (also bundled here as chat_template.jinja) — it carries the anti-rumination tool rules the model was evaluated with; serving without it changes agentic behavior.

Tested on vLLM 0.28, including alongside MTP speculative decoding and on modelopt NVFP4 bases. The released adapter carries vLLM-native (language_model.model.layers.*) key layout — it attaches correctly to the Qwen3_5ForConditionalGeneration module tree. (Plain PEFT exports from a text-only AutoModelForCausalLM run silently attach zero weights in vLLM — basename checks pass, nothing is applied. If you re-export this adapter yourself, keep the VL-layout keys.)

CLI:

vllm serve nvidia/Qwen3.8-27B-NVFP4 \
  --enable-lora \
  --lora-modules cot-lora=/path/to/bf16 \
  --max-lora-rank 16 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

Then request with model: "cot-lora" (adapted) or the base model id (unadapted) — both are live on the same server:

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")
resp = client.chat.completions.create(
    model="cot-lora",  # terse reasoning
    messages=[{"role": "user", "content": "Write a Python lru_cache."}],
    temperature=0.6, top_p=0.95,
)
print(resp.usage.completion_tokens_details.reasoning_tokens)

Hot-loading on an already-running server also works: POST /v1/load_lora_adapter {"lora_name": "cot-lora", "lora_path": "..."}.

Other bases, quants, formats, backends

The adapter is a behavioral edit to the shared Qwen3.8-27B text stack (all attention + MLP + gated-deltanet linear_attn projections; no vision, no MTP weights), so it should stack fine on other quantizations (FP8/INT4/GGUF bases), full-precision bases, and other backends (SGLang, TabbyAPI/EXL3, llama.cpp) — the preference it encodes is not quantization-specific. That said, only the NVFP4 + vLLM combination above has been measured; treat other combinations as untested and validate before relying on them.

Training (summary)

• Round 8 data: a 301-problem pool (HumanEval + MBPP-sanitized, plus mined failures from the round-7 EvalPlus runs) with two pair channels: entropy-pruned (short+pass vs long+pass) and hard-negative (long+pass vs short+fail) — 70 stock pairs + 430 Signal pairs, min-capped to 122 training rows.
• Round 8 method: DPO (trl), β 0.05, lr 1e-5, 3 epochs, effective batch 8, initialized from the round-7 adapter. Round 7's epoch-3 checkpoint overfit (reward accuracy 1.0, held-out collapse after merge), so round 8 is a checkpoint-selection retrain of the same run: per-epoch adapter checkpoints were evaluated independently on the NVIDIA base (held-out-40 + MBPP+), and epoch 1 was selected — the earliest checkpoint before the overfit regime. It is the least-trained checkpoint that still clears the quality gates, which is why its deltas are the smallest of any round.
• Earlier rounds: see archive/r7/README.md and archive/r6/README.md for their training summaries.

Caveats

• Targeted at coding tasks with thinking enabled. Behavioral LoRA, not a knowledge edit.
• The preference is "shorter reasoning, identical answer" — if a task needs long derivation, raise reasoning_effort as usual.
• If you serve with speculative decoding, make sure the generation config has no min_p — vLLM 0.28 rejects min_p under spec decode.

Harness recommendations

How you drive the model matters as much as which build you run. Measured on Terminal-Bench 4.0 (Sep 2026, local vLLM serving): prompt-level guardrails did not reduce token burn or stop grinding — the dominant cost was the harness re-sending the full transcript every step. Zero tool-call loops observed in any trajectory.

• Let the server own sampling — omit temperature/top_p/top_k; the checkpoint config applies temp 0.6 / top_k 20 / top_p 0.95 / repetition_penalty 1.05. Never send min_p (vLLM rejects it under MTP spec decode).
• Serve with the bundled chat template (chat_template.jinja; also at Shockem/froggeric-terse-coder). The anti-rumination rules inject even alongside custom system prompts — bypassing the template (bare completions API, client-side template) is the "repeats already-done steps" failure mode.
• Per-turn max_tokens 4–8k is generous (routine coding answers are 25–150 tokens; thinks median ~38). Use 49152 only as a session floor for agents, not a per-turn target.
• Context management is the biggest lever. Keep ~3 recent tool results, cap older blobs at ~4k chars. Unbounded agent history was worth ~270k+ input tokens per task in our probe — that is a harness property, not a model property.
• Think text arrives in reasoning_content — empty content with non-empty reasoning_content is a normal turn, not an error.
• Raise reasoning_effort for genuinely hard problems — the model still scales deliberation up when the problem needs it (GPQA median ~900 reasoning tokens); terseness targets waste, not deliberation.

Attributions & licenses

This adapter is trained against, and licensed for use with, Qwen/Qwen3.8-27B, © Qwen Team, Alibaba Cloud, Apache 2.0; this adapter is likewise Apache 2.0 and the upstream license and copyright notices are retained. Credits:

• Qwen Team (Alibaba Cloud) — the Qwen3.8-27B base model (Apache 2.0).
• agentionai — Signal-3.8-27B, one of the trace-generation policies.
• p-e-w — the Heretic tool; a heretic-ara abliterated variant of Signal was another trace-generation policy.
• NVIDIA — Qwen3.8-27B-NVFP4 (third policy and eval baseline) and TensorRT Model Optimizer 0.45 (Apache 2.0, quant tooling).
• OpenAI — HumanEval (MIT), and Google — MBPP (CC-BY 4.0): prompt sources for training and held-out evaluation.
• Hugging Face TRL (Apache 2.0) — the DPO trainer; llama.cpp (MIT) — GGUF adapter conversion; Datacurve — DeepSWE, independent evaluation only.