ManniX PRO

ManniX-ITA

32 6 28

https://github.com/mann1x

mann1x

AI & ML interests

None yet

Recent Activity

updated a model about 8 hours ago

ManniX-ITA/Qwen3.5-4B-M8-GGUF

updated a model about 8 hours ago

ManniX-ITA/Qwen3.5-4B-MicroCoder-GGUF

updated a model about 8 hours ago

ManniX-ITA/Qwen3.6-27B-Omnimerge-v4-MTP-GGUF

View all activity

Organizations

None yet

posted an update 18 days ago

Post

227

---
🚀 Gemma-4-A4B 98e v7-coder cohort — loop-fixed re-release. Two 20.8B MoE coders (4B-active), fresh-map prunes of Gemma 4 26B-A4B, 30/128 experts dropped per layer. The headline isn't a benchmark: the agentic loop is
gone at the weights, not papered over by the sampler.

🔧 How: at prune time we force-keep the 46 agentic_eog experts a loop-protection signal flags as load-bearing for clean multi-turn termination (+ shared-FFN α=1.2). Result: 0 loops across 48 seeds on every published
tier.

📊 Q6_K · llama.cpp · greedy · same host (from summary.json):

⚖️ v7-coder (fkbroad code3/lcb2) — balanced coder: LCB-med-55 98.18, HumanEval 98.17, HE+ 92.07, AIME 80.0, MATH-500 95.0, GSM8K 91, IFEval 92, MultiPL-E 89.7, ARC 92.2.

⚡ v7-coderx (code4/lcb3) — code-maximal: all-hard LCB-77 85.71 (cohort-best; 128e 79.22, v7-coder 84.42), HE+ 93.29, GSM8K 93, MATH-500 95.0, AIME 76.67. Whole budget on code.

🎯 Both land near GPQA ~51 — graduate science is the budget axis, neither is a science model. Pick v7-coder for the broad LCB-medium + HumanEval lead; v7-coderx for the all-hard slice and HE+.

🧪 The harness we used to prove the fix is now an omk tool: agentic-loop-harness replays a frozen agentic conversation across a sampler×seed matrix and reports a fail-rate per chat-template, so you can isolate a loop
to one variable. Model-agnostic — any OpenAI-compatible server. The version we shared with Google: google/gemma-4-12B-it#41

📦 Each ships bf16 · GGUF (+ CD-* + imatrix + mmproj vision) · NVFP4A16 (~13 GB) · Ollama.
🔗 ManniX-ITA/gemma-4-A4B-98e-v7-coder-it (+ -it-GGUF, -NVFP4A16) · https://ollama.com/mannix/gemma4-98e-v7-coder
🔗 ManniX-ITA/gemma-4-A4B-98e-v7-coderx-it (+ -it-GGUF, -NVFP4A16) · https://ollama.com/mannix/gemma4-98e-v7-coderx
🔧 https://github.com/mann1x/omnimergekit/tree/main/tools/agentic-loop-harness

posted an update about 1 month ago

Post

319

🚀 Gemma-4-A4B 98e v7-coder cohort — two 20.8B MoE coders (4B-active), fresh-map prunes of Gemma 4 26B-A4B. 30/128 experts dropped per layer from the rebuilt v7 competence maps (audited producers, 10 classes), generic-code 3× + LiveCodeBench 2× on a [24,40] floor, plus the mandatory shared-FFN α=1.2.

📊 Q6_K, llama.cpp, greedy, same host (from summary.json):

🧪 v7-coder — science-augmented (+ targeted_gpqa 1.5). A coder that kept ALL its science: GPQA-D 70.71 (+9.6pp over v6-coder, at PARITY with the unpruned 128e's 67.17), AIME 76.67, MATH-500 92.0, GSM8K 93.0, HE 98.78, HE+ 92.68, LCB-55 96.36, LCB-100 97.0, MultiPL-E 88.67, IFEval 95, ARC 94.8. Edges 128e on GPQA/AIME/GSM8K with no code regression — science recovered by a dedicated targeted_gpqa calibration class.

⚡ v7-coderx — code-maximal (no science term). The strongest coder in the cohort: LCB-med-55 98.18 + LCB-100 99.0 (highest of any Gemma-4 prune to date, +1.8/+2.0pp past the unpruned 128e), MultiPL-E 90.0, HE+ 92.68, HE 95.73, IFEval 95, GSM8K 91, MATH-500 89, AIME 70.0. Whole budget on code; the trade is graduate science (GPQA 48.48).

🎯 v7-coder for strong code + graduate science; v7-coderx for max code (~1.8pp LCB-55 buys ~+22pp GPQA between them). Successor to v6-coder, whose code profile led the 14–22B band (+9pp HE over Qwen2.5-Coder-14B, same-rig).

📦 Each ships bf16 · GGUF (29 tiers + ContribDynamic CD-* + F16 + imatrix + mmproj vision) · NVFP4A16 (native vLLM, ~13 GB) · Ollama (29 tiers + vision-* + :latest=Q4_K_M).

🔗 v7-coder: ManniX-ITA/gemma-4-A4B-98e-v7-coder-it (+ -it-GGUF, -NVFP4A16) · https://ollama.com/mannix/gemma4-98e-v7-coder
🔗 v7-coderx: ManniX-ITA/gemma-4-A4B-98e-v7-coderx-it (+ -it-GGUF, -NVFP4A16) · https://ollama.com/mannix/gemma4-98e-v7-coderx

posted an update about 1 month ago

Post

301

Anthropic liked my agentic council idea and made their own implementation, Dynamic Workflows.
Lovely, dynamic critic role, mine is static, that can be driven by Claude. Done some upgrades on top.
Got my weekly limit reset from 82% to zero but only 50% used.
Been a day and the API backend is still broken, all sessions in retry loops; so sad, so unreliable.
Almost 9h to fix some tests and cut this new release instead of 30m!

🚀 claude-hooks v1.13.0 — the /consultants council learned to argue with itself

The multi-agent council (planner → researcher w/ project tools + citation linting → critic → synthesizer) was missing one thing: a real adversary.

🛡️ Dynamic adversarial review
A new post-synthesis refuter re-reads the council's answer and tries to break it — not a second opinion, a hostile one. It's a singleton, so it stays safe under x-tier multi-model fan-out (no N×M blow-up). The engine
drives an adversary checkpoint: pauses mid-run, streams the pause over SSE, auto-resumes once the skeptic panel reports. Two revived dials — critic_strictness and adversarial_focus — aim the scrutiny (security? perf?
one specific claim?).

🔁 Workflow-driven council
A committed Workflow driver (consult-with-adversarial-review.mjs) runs the whole loop deterministically: ask → review → skeptic panel → accept | follow-up. The council answers, Claude critiques, and the engine either
marks it accepted or auto-issues a focused follow-up and loops — bounded by a flat max_followups cap. A blocking consult --wait makes it scriptable.

🧩 Also in 1.13
• Throttle-aware proxy retry: jittered backoff, Retry-After, a 90s wall-clock deadline, a cross-session circuit breaker — all live at GET /health, so the proxy stops amplifying Anthropic's edge throttle.
• Per-project config: .claude-hooks/consultants.toml promoted via an override_user_global directive.

🐧 Linux / macOS / Windows. PostgreSQL OR SQLite. Local OR cloud LLMs.
🔗 github.com/mann1x/claude-hooks

replied to wenhuach's post about 1 month ago

sorry for the late reply. Is it fine now?

No problem!
One quantization was done but I don't think the evals are working fine with Gemma-4; the inference engine is missing some fixes or the templates needs to be adapted. The scores are very low, quite unrealistic.
I had to spend days to adapt llama.cpp/vLLM/lm-eval to Gemma-4.

replied to wenhuach's post about 2 months ago

Working on it. Gemma 4 had a regression issue, and it has just been fixed.

ManniX-ITA/gemma-4-A4B-98e-v6-coder-it (NVFP4) is already in the Quantization queue (status: Failed). You do not have permission to re-submit failed models — please contact an administrator.

If you manage to fix it please remove my model from the list of failed quants, I can't resend it

replied to wenhuach's post about 2 months ago

Working on it. Gemma 4 had a regression issue, and it has just been fixed.

Thanks! Please remove/re-issue my model from the queue once fixed, I can't do it 🙏

replied to wenhuach's post about 2 months ago

The quantization of my model, ManniX-ITA/gemma-4-A4B-98e-v6-coder-it, at INT4 and NVFP4 failed and at MXFP4 is still running since yesterday. I guess it's stuck.
I wonder, is there a specific problem with Gemma4 A4B MoE architecture or should I just re-queue them?

replied to wenhuach's post about 2 months ago

Sent my model to quantize and eval this morning but seems the first two quants are still in running since this morning. Are they really running? 8 hours seems a bit long.

reacted to wenhuach's post with ❤️ about 2 months ago

Post

4577

🚀 We provide **free** hardware to quantize models at the [Intel Low Bit Open LLM Leaderboard]( Intel/low_bit_open_llm_leaderboard), currently supporting Pure RTN mode powered by AutoRound

⭐ If you find it useful, please consider starring the AutoRound project on [GitHub](https://github.com/intel/auto-round)!

9 replies

posted an update about 2 months ago

Post

228

🚀 Gemma-4-A4B 98e v6-coder (C6v3lcb) — LCB-targeted code prune of Gemma 4 26B-A4B, 20.8B MoE (4B-active). Same C6 recipe as v5-coder, re-steered specifically at LiveCodeBench-medium — the one code bench pruning hurt most.

Not only keeps the lead on Python and closes the gap to 1-2pp in the other coding languages.

It's actually reasoning better, fixing the under-thinking and over-thinking failures of the full experts router.

All this comes with a cost with only 20b, on top of being very specific to coding; about 3x the thinking tokens in LiveCodeBench but it's good thinking that brings home not only more correct answers but in general a more precise and concise output.

📊 SCORES (Q6_K, llama.cpp, greedy, EVAL_PROTOCOL v3)

HumanEval 98.78 — HumanEval+ 93.29 — LCB-medium-55 v4 96.36
LCB-medium-100 96.00 — MultiPL-E macro 88.00 (Rust/Java/JS)
MATH-500 91.00 — GPQA-D 67.17 — AIME 63.33 — IFEval 92.00
vs v5-coder: +10.91 LCB-medium / +7.0 MultiPL-E / +10 AIME, HE+ tie

LCB targeting closed the −9.10pp hole and pushed +1.81pp past the unpruned 128e. Top of the 14–22B coder band: +9.2pp HE over Qwen2.5-Coder-14B-Instruct (89.6 → 98.78).

📦 GGUF SWEEP (all imatrix; Q4_K_M plain — imatrix hurt it)

Q6_K — 17.81 GB — 93.29% (cohort top)
Q3_K_M — 10.51 GB — 92.68% ⭐ value leader (imatrix lifted the 3-bit tiers hard)
IQ4_XS — 11.01 GB — 92.07% ⭐ safe 4-bit
IQ3_XS — 9.22 GB — 92.07% — smallest on the plateau
IQ2_S — 7.83 GB — 89.02% — sub-8 GB code-grade

⚔️ SAME-RIG vs Qwen2.5-Coder-14B (RTX 3090, greedy)

Iso-disk 10.5 GB: Q3_K_M 92.68 vs Qwen Q5_K_M 83.54 → +9.14pp at the same file size
LCB-medium-55 v4, identical split: 96.36 vs 18.18

bf16:
ManniX-ITA/gemma-4-A4B-98e-v6-coder-it ( ManniX-ITA/gemma-4-A4B-98e-v6-coder-it)
GGUF:
ManniX-ITA/gemma-4-A4B-98e-v6-coder-it-GGUF ( ManniX-ITA/gemma-4-A4B-98e-v6-coder-it-GGUF)
Ollama:
https://ollama.com/mannix/gemma4-98e-v6-coder

posted an update about 2 months ago

Post

275

🚀 Gemma-4-A4B 98e v5-coder — code-leaning 20.8B MoE (4B-active), C6 layer-relevance-weighted prune of Gemma 4 26B-A4B. Best 20B-class coder I've shipped.

📊 SCORES (NVFP4A16, vLLM 0.20.2, greedy, EVAL_PROTOCOL v3)

HumanEval 98.17 — HumanEval+ 92.68 — LCB-medium-55 v4 85.45
MATH-500 92.00 — GPQA-D 68.69 — IFEval 94.00
vs v4: +1.22 HE / +1.22 HE+ / +7.27 LCB-medium

Top of the 14–22B coder band: +8.6pp HE over Qwen2.5-Coder-14B-Instruct (89.6 → 98.17). HE+ sanity-audited — no memorization, no silent-empty.

📦 EXTENSIVE GGUF SWEEP (16 plain + IQ tiers + 5 CD recipes, all imatrix-calibrated)

Q8_0 — 21.16 GB — 93.90% (cohort top)
Q4_K_S — 12.21 GB — 93.29% ⭐ plain sweet spot
IQ4_XS — 11.01 GB — 93.29% ⭐ sub-12 GB top

⭐ TWO EXCELLENT SUB-10 GB CONTRIBDYNAMIC CD PICKS (per-layer + IQ-codebook overrides)

CD-IQ4_K_M (Canary W) — 10.29 GB — 92.07% — recommended sub-11 GB
CD-IQ3_XS_L — 9.27 GB — 90.24% — smallest viable code-grade

⚔️ SAME-RIG vs Qwen2.5-Coder-14B-Instruct (RTX 3090, greedy HE+)

11 GB band: v5-coder IQ4_XS wins +9.75pp at -1.49 bpw
12 GB band: Q4_K_S wins +8.53pp
8 GB band: IQ2_S wins +0.61pp at lower bpw

bf16:
ManniX-ITA/gemma-4-A4B-98e-v5-coder-it

GGUF:
ManniX-ITA/gemma-4-A4B-98e-v5-coder-it-GGUF

NVFP4A16:
ManniX-ITA/gemma-4-A4B-98e-v5-coder-NVFP4A16

Ollama:
https://ollama.com/mannix/gemma4-98e-v5-coder

———

🆕 BONUS — Qwen3.6-27B-Omnimerge-v4-MTP-GGUF

Same v4 weights with the native MTP head retained for llama.cpp speculative decoding (PR #22673, --spec-type draft-mtp). 7 imatrix tiers Q8_0 → IQ2_M.

HumanEval: 2.0x decode tok/s
MBPP: 2.33x decode tok/s
Both at +1-2pp pass@1 vs the non-MTP build. GPQA Diamond comparison in flight.

MTP-GGUF:
ManniX-ITA/Qwen3.6-27B-Omnimerge-v4-MTP-GGUF

replied to their post about 2 months ago

The v1.1.0 post had a single "one model wins all" PROD line; v1.8 has per-axis baselines from three live cohorts (May 16-17 '26).

✍️ /consultants coder — per-language routing (mlang v1.0.1, suite hash ddef8095, 6 langs × 13 questions).

Lang	Primary	Fallback
c	glm-5.1:cloud	deepseek-v4-pro:cloud
cpp	deepseek-v4-flash:cloud	kimi-k2.6:cloud
csharp	deepseek-v4-pro:cloud	kimi-k2.6:cloud
go	kimi-k2.6:cloud	deepseek-v4-pro:cloud
python	glm-5.1:cloud	kimi-k2.6:cloud
rust	deepseek-v4-flash:cloud	deepseek-v4-pro:cloud

No single model wins across languages. Out-of-cohort langs (ts/java/ruby/swift/shell) fall to global default glm-5.1:cloud → kimi-k2.6:cloud.

🛠 tool_executor — 48 trials, suite 7921555c:
• gemma4:31b-cloud — 87.5% / Q=5.00 ← winner (tiebreak: quality → wall → call count)
• kimi-k2.6 / deepseek-v4-pro / glm-5.1 — all 87.5% but lost tiebreakers
• gemini-3-flash-preview — 75% (qualifies)
• qwen3-coder-next — 62.5%, DISQUALIFIED. The Python coder winner is NOT a good tool_executor — "best at writing code" ≠ "best at mechanical tool chains for reading code".

⏱ Stall thresholds (M11a, suite c8306c62):
• kimi-k2.6:cloud cold TTFT = 150s → needs stall=390s. The global 300s default would falsely STARTUP_STALL it.
• qwen3-coder-next fastest startup (390ms).
• deepseek-v4-flash slowest p99 wall (255s) → hard_cap raised to 780s.

🧮 Rubric: pass_rate ≥ 70% AND avg_quality ≥ 3.5. Baselines append-only at docs/consultants-skill-eval-baselines.md. Re-bench any new model with claude-consultants skill-eval <suite> --live --accept-cost.

🔗 Bench dirs: benchmarks/consultants/results/2026-05-17/

posted an update about 2 months ago

Post

301

v1.1.0 was Claude + Ollama chat. Eight releases later the stack is a grounded research pipeline plus a local-first memory layer; the token crunch is operational now, not a quality wall.

🚀 claude-hooks v1.8.3 — highlights since v1.1.0.

🧠 /consultants v2 — agentic council, matured.
🛠 tool_executor — PLAN→REPORT lane runs read_file / grep / glob over the codebase before the researcher speaks; claims grounded in tool output, not vibes.
✍️─ coder — sandboxed write_file role with per-language model routing (50KB/file, 1MB/lane caps).
🛡️ CitationLinter — three-layer verifier at the researcher boundary; every path:line claim checked against an mtime-cached code_graph. Catches fabricated filenames before they launder through critics + synthesizer.

💾 M14 cross-session memory (default on).
LangGraph BaseStore wired across four namespaces: research / tool_results / project / user. Per-namespace TTL: research=30d, tool_results=24h, project+user=forever. Hourly Caliber-style distillation reaper summarizes
expiring research into the durable project namespace BEFORE deletion — episodic → semantic, like human consolidation. Originals only dropped after a successful summary write.

🔁 sqlite_vec — full pgvector parity (v1.7).
Hybrid recall via RRF over vector cosine + BM25 (FTS5). KG surface: kg_create_entities / kg_add_observations / kg_create_relations / kg_search_nodes. Bundled sqlite-vec-mcp launcher went 3→8 tools so Cursor / Codex /
OpenWebUI / Claude Desktop share the same .db. Lazy schema migration carries v1.6.x dbs in place, non-destructive.

🧩 llamafile chat + embed (v1.4 + v1.5).
HyDE / reflect / consolidate / get-advice / consultants route to a daemon-supervised local llamafile via the llamafile:// model prefix. Multi-instance LRU, per-label idle reap, sticky CPU fallback. Stack runs
offline now.

🐧 Linux / macOS / Windows. PostgreSQL OR SQLite. Local OR cloud LLMs.

🔗 github.com/mann1x/claude-hooks

1 reply

reacted to Knowurknot's post with 🔥 about 2 months ago

Post

204

Knowurknot/git-hologram git nexus is the one you probably really want to see, but I can't link to it here, so this will have to work. this is cool though, just not really showing off the CAPT with this demo

reacted to danielhanchen's post with 🤝 2 months ago

Post

7784

We collaborated with NVIDIA to teach you how we made LLM training ~25% faster! 🚀

Learn how 3 optimizations help your home GPU train models faster:
1. Packed-sequence metadata caching
2. Double-buffered checkpoint reloads
3. Faster MoE routing

Guide: https://unsloth.ai/blog/nvidia-collab
GitHub: https://github.com/unslothai/unsloth

posted an update 2 months ago

Post

1542

After the Feb/Mar '26 collapse of Claude Code I started building my own framework. The token crunch is still only mitigated, but the reasoning and quality are back — better than before. For research and LLM
training recipes, though, diverse knowledge and a second point of view are crucial. Pairing my Claude Max sub with an Ollama Pro sub has already saved me from days of botched trainings — multiple frontier
models helping Claude is next level. Acting as the middleman myself was interesting but inefficient, so I shipped skills that let Claude talk to Ollama models directly.

🚀 claude-hooks v1.1.0 ships two LLM-to-LLM skills.

💬 /get-advice — single-shot second opinion. Claude runs a multi-turn conversation with a configured Ollama advisor; the advisor grounds in your project through read_file / grep / glob / list_files /
recall_memory tools. Effort tiers cap fresh-session retries.

🤝 /consultants — multi-agent council for cross-cutting questions:
🧩 planner → researcher → critic → synthesizer

Each role runs its own Ollama model. 💾 Sessions persist to disk (summary.md + transcript.md + SQLite per-role message threads); closed sessions reopen and produce follow-ups 🔁 indistinguishable from warm
ones.

🎯 x-tier effort multiplies diversity:
• xmedium / xhigh — researcher fans across N models in parallel
• xmax — + multi-critic + meta-critic combine; critics anonymized as "Critic 1/2/3" to avoid model-bias

🛡️ Cloud-flap recovery, three layers: 15-attempt / ~15min retry budget; synthesizer failure-fallback model chain; degraded-answer composer surfaces researcher + critic work even when synthesis fails.

📊 7 cloud models benchmarked & Claude-graded on locked queries:
• PROD-READY (P:A R:A C:A S:A): kimi-k2.6, gemma4:31b, glm-5.1
• Role specialists: minimax-m2.7 (critic), qwen3.5 (planner)

🐧 Linux/macOS/Windows. No per-project setup.

🔗 github.com/mann1x/claude-hooks

reacted to mlabonne's post with 👍 2 months ago

Post

4015

Big update to llm-datasets, my curated list of datasets and tools for post-training LLMs.

> Added many new datasets
> New "thinking" column
> Refreshed recommended tools.

Thanks to everyone who told me they used it for their research at ICLR, you motivated this update!

2 replies

reacted to DedeProGames's post with 🔥 2 months ago

Post

8591

GRaPE 2 Pro is now available.

SL-AI/GRaPE-2-Pro

This is the flagship model of the GRaPE 2 family and the largest model I have trained to date, sitting at 27B parameters. It is built on Qwen3.5-27B and trained on a closed-source proprietary dataset, with roughly half of post-training focused on code and the rest split between STEAM subjects and structured logical reasoning. It punches seriously above its weight class.

GRaPE 2 Pro supports multimodal input (image + text) and features 6 thinking modes via the tag. This gives you real control over how hard the model thinks, from skipping the reasoning phase entirely with minimal, all the way up to xtra-Hi for deep, extended thought on hard problems. For most agentic use, auto or low is the move to keep things snappy.

It also runs on consumer hardware. You can get it going with as low as 12GB of VRAM on a quantized build.

If you want to try it out and give feedback, that would be really appreciated. Email us at contact@skinnertopia.com

1 reply

posted an update 2 months ago

Post

259

🚀 Exciting week, 2 new research projects and 2 new tools!

▶ Mythic-RDT - OpenMythos blueprint with a retrofit-recurrence fine-tune
https://github.com/mann1x/Mythic-RDT

▶ cross-tokenizer-distill (CTD) - knowledge distillation across different tokenizer vocabularies
https://github.com/mann1x/cross-tokenizer-distill

For Mythic-RDT, I have chosen the pretty outdated DS-Coder-V2 16B.
It's small enough to not need more than 48GB VRAM but once I leaned on KL for depth recurring fine-tune (couldn't go above parity to T=1 with T=4, not the best for 4x inference time), started investigating the KL recipe and questioned the teacher, same DS-Coder-V2 but at BF16.
For a better teacher the option would have been just one, DS-Coder-V2-236B. Not only so big that I'd need 4xH100 to run but also surpassed even by Qwen3-Coder-32B on HE/MBPP.
Hence here's CTD tool, validated but still in development to find a good recipe for Qwen->DS distill.

▶ Qwen3.5-4B-MicroCoder - code-leaning and reasoning merge of Qwen3.5-4B
ManniX-ITA/Qwen3.5-4B-MicroCoder

▶ Omnimergekit - merge toolkit, merge & quantization scripts, experiments logs
https://github.com/mann1x/omnimergekit

You can find my merge toolkit and scripts in the repo, so they don't get scattered over the HF repos.
Interesting experiment with MicroCoder; only a couple of base, reasoning broken, coding fine-tune to merge with the excellent instruct reasoning JackRong-v2.
The result is not truly exciting but manages to improve LiveCodeBench above JR-v2, improve MBPP and not completely breaking reasoning.
This is achieved with omnimergekit using differential signals generated by the delta vs the base model from the good and wrong answers delta between the sources (HE/MBPP/AIME).
The very long eval sessions proved that the method does not just bias the scores of these evals but improve others even above the baseline.

replied to their post 2 months ago

🔄 Follow-up to the M1..M5 study above — re-ran M4 against the updated PR #682 head ("turbo" branch by @Tusm11, HEAD 8d989f6) with rebalanced hyperparams (equal weights w=1/1, density 0.7, closer to the PR's worked example). Same AttnLRP signal as M4-orig, same sources.

▶ Qwen3.5-4B-M4-v2-ex-LRP-turbo
ManniX-ITA/Qwen3.5-4B-M4-v2-ex-LRP-turbo

Q6_K HE / MBPP pass@1, M4-v2 inserted:

M1 Vanilla DARE-TIES → 51.22 / 47.00
M2 OMv2 (no signal) → 52.44 / 49.40
M3 OMv2 + Fisher → 57.93 🥇 / 48.80
M4 ex-LRP (PR #682 orig) → 51.22 / 49.40
M4-v2 ex-LRP (PR #682 turbo) → 55.49 / 52.20 🥇
M5 OMv2 + LRP → 53.05 / 51.40

Δ M4-v2 vs M4-orig: +4.27 pp HE, +2.80 pp MBPP. M4-v2 takes the MBPP medal of the whole study (overtakes M5) while staying competitive on HumanEval. The turbo code path + rebalanced hyperparams clearly beat the original PR head on this configuration.

Findings refresh: Fisher still leads HE; ex-LRP (turbo) now leads MBPP, narrowly ahead of OMv2+LRP. Both LRP variants land within 1 pp on MBPP — strong signal that LRP-driven sparsification is doing real work for code-gen on small Qwen merges.

Big thanks to @Tusm11 for the supercharged ex-LRP turbo head — multimodal support + Iron-Man stabilization + in-place math are a real upgrade. Posted full results + the 6 patches needed to run it against Qwen3_5ForConditionalGeneration on the PR thread:
https://github.com/arcee-ai/mergekit/pull/682

ManniX PRO

AI & ML interests

Recent Activity

Organizations

ManniX-ITA's activity