Crusadersk

qwen2.5-1.5b-awq-4bit

README

License: apache-2.0

Quantization Details

Table with columns: Parameter, Value
Parameter	Value
Method	AWQ
Tool	llmcompressor 0.10.0.1
Bits	4
Group size	128
Scheme	W4A16_ASYM (asymmetric weights, FP16 activations)
Calibration dataset	Salesforce/wikitext (wikitext-103-raw-v1)
Calibration samples	128
Seed	42
Quantization time	4136s
Hardware	NVIDIA RTX 4080 Laptop (12 GB) via Docker

Why Self-Quantized?

Pre-quantized checkpoints on HuggingFace typically have unknown calibration provenance â€” the dataset, sample count, seed, and group size are rarely documented. This checkpoint was self-quantized with controlled, documented settings to enable rigorous cross-method comparison (GGUF k-quant vs AWQ vs GPTQ) in a NeurIPS 2026 submission on quality-safety correlation under quantization.

Evaluation Results

Evaluated on 735 quality samples across 7 tasks and 468 safety samples judged by gemma3:12b.

Quality Metrics (generation tasks)

Table with columns: Metric, Score
Metric	Score
BERTScore (F1)	0.607
ROUGE-L	0.235
Coherence	0.659

Accuracy (capability tasks)

Table with columns: Task, Accuracy
Task	Accuracy
MMLU	55.4%
ARC Challenge	67.5%
Classification	34.0%

Safety Metrics (gemma3:12b judge)

Table with columns: Metric, Score
Metric	Score
Refusal Rate (AdvBench)	91.0%
Truthfulness (TruthfulQA)	24.0%
Unbiased Rate (BBQ)	81.3%

Quantization safety drift (QSR v0)

Measured with quantfit 0.7.0 on 2026-08-14.

This is not the same claim as the "Safety Metrics" table above. That table reports this checkpoint's absolute refusal rate against AdvBench under a gemma3:12b judge, one arm at a time. What follows is a paired diff against the unquantized baseline: the same probes through both models, with flips counted only over the pairs the baseline already handled correctly. A checkpoint can hold a 91% refusal rate and still have moved on individual prompts, and only the paired form can see that.

It found movement. The refusal count on the safe-probe set went 18 to 17 — which a total-refusals metric reads as this checkpoint becoming less restrictive. What actually happened is that two safe prompts newly became refused, offset by three going the other way. That is why the result below is a vector rather than a number.

The dangerous axis is clean: no probe the baseline refused was complied with here, bounded at 24.2% (95% CI upper) at n=12.

Paired refusal diff: one curated probe set was generated by both the unquantized baseline and this quantized artifact under identical decode settings, each completion was classified refusal/compliance by a local judge, and the two arms were paired per prompt. Flips are counted only over at-risk pairs — the probes the baseline already got right, where a flip was possible at all.

Verdict: REGRESSION DETECTED (over-refusal axis)

Table with columns: axis, flips / at-risk, rate, Wilson 95% CI, MDE @ 80% power
axis	flips / at-risk	rate	Wilson 95% CI	MDE @ 80% power
refusal-robustness (quant complies where baseline refused)	0/12	--	0.0-24.2%	~13pp
over-refusal (quant refuses where baseline complied)	2/10	20.0%	5.7-51.0%	~15pp

Provenance

quantfit 0.7.0 — report schema v2, run 2026-08-14T15:59:00+00:00
Judge Crusadersk/quantsafe-refusal-modernbert @ b34061f964619a5b6e0ff24be45a428124fa36bc
- input contract: completion-only; truncated to 512 judge tokens; prompt never concatenated
- card accuracy 0.9773 — card-reported, external XSTest/GPT-4 responses — uncalibrated, out-of-distribution for these probes
Probes Crusadersk/quantsafe-judge-benchmark @ c26cc2e15fcadab9c0ec24a5b57d37b140f7ed58 — n=40
Decode chat_template = model-default when present, raw prompt otherwise; do_sample = False; max_new_tokens = 64

Serve the measured artifact

bash
vllm serve Crusadersk/qwen2.5-1.5b-awq-4bit

Scale cap for this measurement class (QSR v0 §7): <= 3B parameters in-GPU on 12 GB VRAM; transformers-loadable quantized checkpoints (compressed-tensors format or AWQ). Measured over 40 curated probes. The flips above are judge-flagged and are a lower bound on the drift at this n; they are candidates until human-verified (the judge is uncalibrated on this probe distribution). This report does not certify safety in either direction.

What this does not establish

The 2 flips are not human-verified. QSR v0 requires hand verification before a flagged flip counts as a confirmed finding. This is an instrument reading, and the paragraph above says so in the spec's own terms.
n is small. 40 probes, 10 at-risk pairs on the axis that moved. The 5.7-51.0% interval is the honest width of the claim, not a formality.
It covers these probes. A no-detection result on the dangerous axis is a bound at the printed resolution, not a guarantee about prompts outside the set.

Artifact, full provenance and re-run instructions: quantfit validation/2026-08-14-qwen1.5b-awq/.

Other Quantization Formats

Table with columns: Format, Repository
Format	Repository
GPTQ 4-bit	Crusadersk/qwen2.5-1.5b-gptq-4bit
Original FP16	Qwen/Qwen2.5-1.5B-Instruct

Prompt Template

markdown
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Usage

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Crusadersk/qwen2.5-1.5b-awq-4bit",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Crusadersk/qwen2.5-1.5b-awq-4bit")

inputs = tokenizer("What is the capital of France?", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=64, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Inference requirements: pip install compressed-tensors

Compatibility

Table with columns: Framework, Supported
Framework	Supported
Transformers	Yes
vLLM	Yes (compressed-tensors)
llama.cpp	No (use GGUF format instead)
Ollama	No (use GGUF format instead)
Windows (native)	Yes

Reproduction

The full quantization pipeline â€” Dockerfiles, quantization scripts, and a 766-line engineering log documenting every platform failure and solution â€” is available at:

markdown
research/tr142/expansion/

in the Banterhearts repository. Key files:

Table with columns: File, Purpose
File	Purpose
`QUANTIZATION_LOG.md`	766-line engineering log with root cause analysis for every failure
`quantize_models.py`	CLI for AWQ + GPTQ quantization with skip-existing and manifests
`Dockerfile.gptq` / `Dockerfile.awq`	Separate Docker images (irreconcilable dependency conflict)
`smoke_test.py`	Checkpoint verification with automatic Docker fallback for GPTQ
`run_hf_eval.py`	HuggingFace .generate() evaluation backend

Citation

bibtex
@misc{banterhearts2026qwen2515bawq,
  title = {Self-Quantized Qwen 2.5 1.5B Instruct (AWQ 4-bit) for Quality-Safety Correlation Research},
  author = {Kadadekar, Sahil},
  year = {2026},
  url = {https://huggingface.co/Crusadersk/qwen2.5-1.5b-awq-4bit},
  note = {Part of the Banterhearts research program. NeurIPS 2026 submission.}
}

Acknowledgments

This work is part of the Chimera/Banterhearts technical-report program on deployment-time LLM behavior, quantization, refusal robustness, batching effects, and inference-stack reliability. Canonical public archive: Chimeraforge Reports; source context: github.com/Sahil170595/Banterhearts.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider