Jakubrd4/bielik-q2-sharp

Bielik Q2# Research: Evaluation Results & Documentation

Evaluation results, scripts, logs, and reports from 2-bit quantization research on speakleash/Bielik-11B-v2.3-Instruct.

Variants

Variant A: QuIP# E8P12 (successful)

• Method: QuIP# with E8P12 lattice codebook, 2-bit
• Model: Jakubrd4/Bielik-11B-v2.3-Instruct-QuIP-2bit
• Size: 3.26 GB (vs ~22 GB FP16, ~6.7x compression)
• Normalized avg (22 tasks): 61.10 (vs 65.71 FP16, ~93% retention)
• Evaluation: Full Polish LLM Leaderboard (22/23 tasks, eq_bench excluded)

Variant B: SpinQuant + GPTQ (unsuccessful)

• Method: SpinQuant rotation matrices (R1 + R2) + GPTQ 2-bit quantization
• Result: Model produced incoherent output after quantization
• Partial eval: DYK multiple choice only (62.88% acc)

Directory Structure

variant_a/
  eval/                              # Full evaluation results (lm-evaluation-harness JSON)
    gen_full_results.json            # Generative tasks (42 MB)
    mc_full_results.json             # Multiple choice tasks (152 MB)
    remaining_full_results.json      # Remaining tasks incl. perplexity (289 MB)
    variant_a_all_results.json       # Combined summary scores
    variant_a_gen_results.json       # Gen summary
    variant_a_mc_results.json        # MC summary
  report/
    variant_a_report.md              # Technical report (Polish)
  logs/
    auto_chain.log                   # Automated eval chain log
    eval_full_mc.log                 # MC evaluation log
    gen_log.txt                      # Generative evaluation log

variant_b/
  eval/
    dyk_mc_results.json              # Partial MC eval (DYK task only)
  config/
    config.json                      # Model config
    quantize_config.json             # GPTQ quantization config
    quantization_meta.json           # Quantization metadata
  report/
    variant_b_summary.md             # Full analysis (Polish)
    variant_b_summary_short.md       # Short summary
  logs/
    pipeline.log                     # SpinQuant + GPTQ pipeline log
    step4b_output.log                # GPTQ quantization logs (3 attempts)
    step4b_v2.log
    step4b_v3.log
    step5_mc.log                     # MC evaluation log
    step5_output.log                 # Generation test log
    full_eval.log                    # Full eval attempt log
  rbin_info.txt                      # Rotation matrix info

scripts/
  eval_polish_quip.py                # QuIP# evaluation wrapper (patched for Mistral)
  full_cloud_eval.sh                 # Cloud GPU setup & full eval script
  run_eval.py                        # MC evaluation runner

Results (22 tasks, normalized scores)

QuIP# E8P12 vs IQ2_XXS — both evaluated on speakleash/Bielik-11B-v2.3-Instruct.

#	Task	Type	QuIP#	IQ2_XXS	FP16	Delta
1	polemo2-in_mc	mc	74.62	75.33	78.42	-0.71
2	polemo2-out_mc	mc	59.64	65.09	66.05	-5.45
3	8tags_mc	mc	70.27	72.46	76.01	-2.19
4	belebele_mc	mc	74.57	75.34	82.59	-0.77
5	dyk_mc	mc	55.88	52.66	58.17	+3.22
6	ppc_mc	mc	61.96	54.73	64.54	+7.23
7	psc_mc	mc	89.19	85.09	85.78	+4.10
8	cbd_mc	mc	14.11	17.85	26.95	-3.74
9	klej_ner_mc	mc	21.46	26.34	28.63	-4.88
10	polqa_rerank_mc	mc	61.79	60.98	67.12	+0.81
11	polemo2-in_g	gen	71.07	69.88	76.52	+1.19
12	polemo2-out_g	gen	55.48	51.63	59.96	+3.84
13	8tags_g	gen	72.64	71.71	74.22	+0.93
14	belebele_g	gen	79.35	77.04	84.28	+2.31
15	dyk_g	gen	48.96	54.86	58.30	-5.90
16	ppc_g	gen	63.68	62.99	64.20	+0.69
17	psc_g	gen	93.67	93.36	84.63	+0.31
18	cbd_g	gen	17.85	28.96	28.06	-11.11
19	klej_ner_g	gen	29.15	29.44	31.44	-0.30
20	polqa_open_g	gen	91.11	88.11	92.19	+3.00
21	polqa_closed_g	gen	73.42	68.22	84.11	+5.19
22	poquad_open_g	gen	64.23	64.33	69.29	-0.10

Summary:

• Normalized avg: QuIP# 61.10 vs IQ2_XXS 61.20 (delta -0.11)
• FP16 retention: QuIP# 93.2% vs IQ2_XXS 93.4%
• Head-to-head: QuIP# wins 11, IQ2_XXS wins 8, ties 3
• eq_bench excluded (private dataset)

Related Resources

• Quantized model (Variant A): Jakubrd4/Bielik-11B-v2.3-Instruct-QuIP-2bit
• Hessian matrices: Jakubrd4/bielik-quip-e8p12 (hessians/ directory)
• Base model: speakleash/Bielik-11B-v2.3-Instruct
• Polish LLM Leaderboard: speakleash/open_pl_llm_leaderboard