base_model: XHToken/Spark-X2.5-4B-Base
base_model_relation: quantized
license: apache-2.0
library_name: gguf
pipeline_tag: text-generation
language:
- en
- zh
tags: - gguf
- llama-cpp
- quantization
- sparkx2_5
- fit-gguf
- uncensored
- abliterated
Spark-X2.5-4B-abliterated · FIT-GGUF
Four fidelity tiers of a 1M-context hybrid-attention 4B — every shipped file re-verified against its own BF16, byte-exact and reproducible.
English · 简体中文 📖
Every file in this repository was planned, executed and verified by FIT-GGUF, an open-source, deterministic tensor-level planning layer on top of llama.cpp quantization. Standard GGUF quantization asks you to pick one of a handful of presets; FIT-GGUF instead asks what quality do you want, then finds and verifies the smallest GGUF that demonstrably meets it.
Traditional GGUF gives you presets. FIT gives you a fidelity contract: macro KL ≤ tier anchor ∧ same-top ≥ model-calibrated floor.
| Deterministic size prediction & byte-exact delivery | ✅ Validated (G2 gate, delta = 0) |
| Universally optimal tensor allocation | ⚠️ Not established — FIT claims verified fidelity contracts, not a universal quality optimum |
The method, the preregistered research record and the fit CLI are open source: github.com/Scorp1o117/FIT-GGUF
The source model is an abliterated, refusal-removed model with no meaningful built-in guardrails, and may comply with harmful, illegal or unsafe requests. Use it only where you can provide appropriate moderation, access control and legal review. Do not deploy it to end users without your own safety layer.
源模型经过拒答方向消融，不具备可靠的内置安全护栏；请仅在合法、受控、具备审核与滥用防护的环境中使用，使用者自行承担部署责任。
The refusal-direction ablation was performed locally with the abliterix pipeline: mean-LoRA steering — rank-8 full-norm LoRA adapters built from per-layer mean refusal directions, with projected abliteration, a gaussian decay kernel over depth, and per-layer vector scope. The exported weights are Optuna trial #615 of study spark_x25_4b_lora_v30, selected from the measured Pareto front.
| Metric | Original | T615 |
|---|---|---|
| Keyword refusals (100 prompts, held-out slice) | 100 | 7 |
| 3-token full-distribution KL vs original (nats/token) | — | 0.1468 |
| Validation KL (mean over held-out prompts) | — | 0.148 |
Honest selection trade-off: the pre-registered "ship bar" was refusals ≤ 10 ∧ KL ≤ 0.05. T615 clears the refusal bar with margin but deliberately trades weight distance for openness (KL 0.147 > 0.05) — the previous export T580 (KL 0.093) left sexual-content soft-refusals and was replaced by user choice. This is an openness/faithfulness Pareto decision, reported as measured.
| Trial | Refusals /100 | KL (nats/token) | Note |
|---|---|---|---|
| 615 | 7 | 0.1468 | this export |
| 389 | 10 | 0.1433 | Pareto neighbor |
| 221 | 3 | 0.2299 | most open |
| 580 | 20 | 0.0929 | previous export (leftover refusals) |
All fidelity-tier measurements in this repository are taken against the T615 BF16 weights themselves — the ablation delta is baked into the reference, so the tier numbers below quantify quantization loss only.
File size ≠ RAM/VRAM usage. KV cache, compute buffers and runtime overhead are separate. The native context is 1M tokens; the hybrid attention (3 sliding-window : 1 full) keeps most depth cheap, but the 9 full-attention layers still need ~34 GiB of f16 KV at 1M tokens — pick a sane -c. Naming: Spark-X2.5-4B-abliterated-FIT-<tier>-<size>-<dominant>.gguf
| Tier | GiB | Dominant | Macro KL ↓ | Same-top ↑ | Gates (KL ≤ / top ≥) | Measured positioning |
|---|---|---|---|---|---|---|
| ⭐ QUALITY | 3.147 | Q6_K | 0.0289 | 93.50% | 0.05 / 93.16% | Max verified quality; the Q6_K preset is itself the minimum verified PASS |
| BALANCED | 2.680 | Q5_K_S | 0.0739 | 89.15% | 0.10 / 89.14% | Best size/quality trade-off; nothing smaller passes both gates |
| COMPACT | 2.432 | Q4_K | 0.1458 | 84.88% | 0.15 / 84.88% | FIT tensor-level recipe — fills the preset gap (Q5_K_S 0.074 → Q4_K_M 0.170) |
| MINI | 2.260 | Q4_K_S | 0.1834 | 82.61% | 0.20 / 82.61% | Smallest verified PASS |
Quick picks
All numbers are protocol-scoped observations (five fixed 64 KiB domains vs this model's aligned BF16), not an application benchmark or a universal ranking. Honest scope: on this model the preset ladder is strong — Quality/Balanced/Mini minimum verified PASS are the native presets themselves; FIT contributes the verification and the Compact gap-fill. All four tiers are same-top-bound: the guard floors, not KL, decide the sizes.
Quality improves monotonically across the healthy preset ladder, and every tier above is same-top-bound — the model-calibrated floors, not KL, are what separate PASS from FAIL here. The 2-bit region is reported as measured: Q2_K/IQ2_* collapse (macro KL 0.8–3.9). The COMPACT tier fills the gap between Q5_K_S (0.074) and Q4_K_M (0.170) where no native preset exists.
These GGUFs use the spark2_5 architecture. You need a llama.cpp build with Spark2_5 support — PR #27868 (open at release time, 2026-09) or any later release that includes it. Upstream master without this PR cannot load these files.
llama.cpp
./llama-server \ -m Spark-X2.5-4B-abliterated-FIT-BALANCED-2.68GiB-Q5_K_S.gguf \ -ngl 99 -c 32768
./llama-cli
-m Spark-X2.5-4B-abliterated-FIT-QUALITY-3.15GiB-Q6_K.gguf
-ngl 99 -c 8192 -p "Hello" -n 256
The chat template is embedded in every file. Once Spark2_5 support merges, any llama.cpp-based runner (llama-cli, llama-server, LM Studio, KoboldCpp, Jan, …) loads these files directly. This is a dense 4B text model — no MTP head, no vision projector.
| Runtime | llama.cpp @ PR #27868 head (ae320b1) · Linux x86_64 · ROCm (gfx1151) — perplexity/KL code byte-identical to the pinned eval-v1 build |
| Command shape | llama-perplexity -ngl 99 -t 16 -c 512 -b 512 --kl-divergence … |
| Reference | This model's own BF16 logits (T615 weights) |
| Domains | wiki_test · wiki_valid · Chinese · code · agent_chat (five fixed 64 KiB slices, macro mean) |
Tier verification: a tier ships only if macro KL ≤ its anchor and same-top ≥ a model-calibrated guard floor (15-point preset ladder + 2 gap probes; floors are exact-model-scoped and are not transferred across models). Every shipped file was re-evaluated as the exact shipped bytes; the COMPACT recipe is reproducible byte-for-byte.
Allocator scope: the balanced v0.1 policy was used as-is (no model-specific refine profile exists for spark2_5). This release claims deterministic size planning and measured, gate-verified fidelity for these specific artifacts — it does not claim a universally optimal allocation.
✅ 4 gate-verified tier GGUFs · ✅ the BF16 reference GGUF (the exact weights every measurement above is taken against) · ✅ chat template embedded in every file · ✅ checksums (SHA256SUMS.txt) · ✅ labeled quality curves (results/) · ✅ tier manifest (FIT-TIERS.md)
❌ 2-bit quantization classes — measured collapse on this model (macro KL 0.8–3.9), excluded by design · ❌ no vision projector / no MTP head (dense text model)
The abliteration (T615) was performed locally from XHToken/Spark-X2.5-4B-Base; this repository contributes the abliterated weights and the FIT quantization plans and artifacts.
sha256sum -c SHA256SUMS.txt
The evaluation slices are public in the FIT-GGUF repository; the calibration ladder, gap probes, search audit and guard profile for this release are retained in the FIT-GGUF experiment records (2026-09-06-spark-x25-4tier). Exact-size behavior is scoped to the recorded source metadata and the recorded llama.cpp build; changing the converter, runtime, source layout or metadata requires revalidation.
Apache-2.0, inherited from the base model — follow the upstream license and model-card requirements.
XHToken / SparkLLM — the Spark-X2.5 model family · llama.cpp PR #27868 (KnightYao) — Spark2_5 support · abliterix — the local ablation pipeline (trial T615) · FIT-GGUF — verified size-exact quantization. FIT-GGUF is an independent project, not affiliated with XHToken, SparkLLM or llama.cpp.