Launch80/Qwopus3.8-27B-Coder-Flash-quality-eval

Qwopus3.8-27B-Coder-Flash: quality-eval reference

This is the measurement data behind the Launch80 INT4 study of Jackrong/Qwopus3.8-27B-Coder-Flash. It contains:

• the BF16 reference logits: top-64 log-probs on a fixed 512 × 2048-token set, plus two perplexity sets
• the exact token ids they were computed on
• per-item pass/fail results for seven tasks
• the scripts that produce and score all of it

It's meant for scoring your own quant of this model on the same harness. You don't need to run the 55 GB BF16 model.

The quantized model is Launch80/Qwopus3.8-27B-Coder-Flash-Int4-baked. The full write-up is the study.

Files

path	what it is
eval-set.json	Token ids for every scored window: kld 512 × 2048, ppl-wiki 40 × 2048, ppl-code 40 × 2048. meta.sha256_ids identifies the set.
reference/bf16-logprobs.npz	BF16, teacher-forced, speculation off. For each set: <set>.target float32 [W, 2047] (log-prob of the true next token), <set>.ids int32 [W, 2047, 64] and <set>.lp float16 [W, 2047, 64] (top-64 ids and log-probs, sorted).
reference/items/<task>.jsonl	BF16 per-item results, one {"id", "correct"} per line.
arms/v1.0/logprobs.npz	The same logits for the Int4-baked checkpoint at tag v1.0. Weights are unchanged through v1.1.1.
arms/v1.0/items/<task>.jsonl	Per-item results for that checkpoint, served as shipped (MTP k=6).
arms/v1.0/kld-vs-bf16.json, arms/v1.0/stats/*.json	kld.py and stats.py output for that arm.
eval-set-code-sources.jsonl	The 298 codeparrot source files behind the code windows of eval-set.json: repo, path and licence for each.
calibration/calib-body.jsonl	The exact calibration token ids the INT4 body was quantized on: 512 rows, 975,908 tokens, sha256 f2011598…. See Calibration.
calibration/code-sources.jsonl	The 146 codeparrot source files behind its code rows: repo, path and licence for each.
scripts/	logprobs.py (dump the logits from any vLLM OpenAI server), kld.py (ppl / KLD / top-1 vs the reference), stats.py (paired McNemar + bootstrap CI per task), build_eval_set.py (how eval-set.json was built), quant_body.py and calib_tools.py (how the calibration set was built, and the GPTQModel body quantization).
SHA256SUMS	Hashes of every file.

Score your own quant

# 1. serve your quant with vLLM: --max-logprobs 64, speculative decoding OFF
# 2. dump its logits on the exact same token ids
uv run --with numpy python scripts/logprobs.py --base http://127.0.0.1:8000 --model <served id> \
    --set eval-set.json --out mine.npz
# 3. compare with BF16
uv run --with numpy python scripts/kld.py --ref reference/bf16-logprobs.npz --test mine.npz --json mine-kld.json

kld.py refuses to compare files built from different eval sets, by checking eval_set_sha256.

For task results, write one {"id": ..., "correct": true|false} line per item, using the same item ids as reference/items/. Then run python3 scripts/stats.py --ref reference/items/<task>.jsonl --test <yours>.jsonl.

Method

• KLD is KL(BF16 ‖ quant) per position, summed over the BF16 top-64 ids plus one tail bucket.
  • A BF16 id missing from the quant's top-64 is given the quant's leftover probability mass, spread evenly over the vocab ids outside its top-64.
  • The BF16 top-64 covers 95.3 % of probability mass on average, which bounds the size of that approximation.
  • Top-1 agreement is an exact argmax match.
  • Positions 1..2047 of each window are scored.
• Perplexity is exp(mean NLL) of the true next token on the same positions.
• Tasks are all greedy and paired item-by-item against BF16:
  • HumanEval+ (164) and MBPP+ (378) with evalplus 0.3.1
  • a LiveCodeBench slice (175)
  • GSM8K (1319), IFEval (541) and an MMLU-Pro subset (1000) with lm-eval 0.4.13
  • a BFCL subset (400) with bfcl-eval 2026.3.23
• Hardware: the BF16 reference ran on an RTX PRO 6000 (vLLM 0.30, CUDA). The v1.0 arm ran on an Intel Arc Pro B65 (vLLM 0.30, XPU). The KLD therefore includes any difference between the two engines, not only the quantization.

Calibration

calibration/calib-body.jsonl holds one JSON list of token ids per line, already tokenized and chat-templated with this model's tokenizer. It was built with seed 608 and is disjoint from every eval set above. The 204 code rows were re-derived from the public source file and match the stored ids exactly.

rows	share	source
0–152	30 %	C4 en validation shard 0, as 2048-token windows (the eval set uses shard 1)
153–356	40 %	codeparrot/codeparrot-clean-valid file 54, the first 90 % of its files, as 2048-token windows (the eval set uses the last 10 %)
357–458	20 %	NousResearch/hermes-function-calling-v1 @ dae3e1d, func-calling.json: 51 conversations with parallel calls and 51 with single calls, rendered with tools= through the model's own chat template, each cut to a window that holds its first assistant tool call
459–511	10 %	HuggingFaceH4/ultrachat_200k test_sft: whole conversations through the chat template, at least 256 tokens, cut at 2048

Rows are 280 to 2048 tokens long. GPTQModel used them as they are, for the v1.x body. Tools that need equal-length rows can concatenate them and re-cut into 2048-token rows, which gives 476 full rows. AutoRound, with batch 8, then keeps whole batches: 472 rows.

v1.0 at a glance

All numbers below were measured against reference/. The harness is this dataset's kld.py and stats.py.

metric	BF16	v1.0 (Int4-baked)
KLD mean (median, p99)	0	0.0412 (0.0140, 0.375)
top-1 agreement	100 %	92.84 %
ppl wikitext-2 / code	5.980 / 2.598	6.082 (+1.72 %) / 2.645 (+1.83 %)
HumanEval+	90.24	88.41 (CI −5.5…+1.8)
MBPP+	76.98	75.66 (CI −3.2…+0.5)
LiveCodeBench slice	41.14	40.57 (CI −5.1…+4.0)
GSM8K	86.96	86.66 (CI −1.2…+0.6)
IFEval	80.22	80.78 (CI −2.2…+3.1)
MMLU-Pro subset	71.80	70.90 (CI −2.7…+0.9)
BFCL subset	89.00	88.25 (CI −3.0…+1.5)

The CIs are paired bootstrap 95 % intervals on the difference, in points. No task difference is significant: McNemar p ≥ 0.27 for every task.

Sources and licences

• Our files (logits, per-item results, scripts) are released under Apache-2.0, matching the model.

• The token ids in eval-set.json encode public text, which keeps its own licence:

  • ppl-wiki: WikiText-2 test split (CC BY-SA 3.0)
  • kld, first 256 windows: C4 en validation shard 1, allenai/c4 (ODC-BY)
  • kld, last 256 windows, and ppl-code: the last 10 % of codeparrot/codeparrot-clean-valid file 54. These are GitHub source files under their own original licences.

• Task items are referred to by id only. No benchmark text is redistributed.

• The token ids in calibration/calib-body.jsonl also encode public text:

  • C4 (ODC-BY)
  • codeparrot-clean-valid: GitHub files under their own licences. Each file and its licence is listed in calibration/code-sources.jsonl.
  • NousResearch/hermes-function-calling-v1 (Apache-2.0)
  • HuggingFaceH4/ultrachat_200k (MIT; its conversations were generated with ChatGPT)

• Code licences by file, from codeparrot's per-file license field:

  set	files	MIT	Apache-2.0	BSD-3	GPL-3.0	GPL-2.0	AGPL-3.0	other
  calibration	146	52	23	13	28	20	3	7 (BSD-2, LGPL-2.1, MPL-2.0, Unlicense)
  eval set	298	79	60	38	45	33	15	28 (BSD-2, LGPL, Unlicense, EPL, MPL, CC0, ISC)

  Some of these are copyleft (GPL, AGPL). The windows are token ids of excerpts from those files, and each excerpt stays under its file's original licence. The manifests name every file so each one can be checked.

Our own speculative-decoding data is not in this dataset.

Citation

If you use this data, link this repo and the study.

Community

Questions, results, or a model you want tested? Join the Launch80 Discord: https://discord.gg/launch80

If this work is useful to you, please support the server: https://launch80.com/supporter