GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
alinakhay
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
alinakhay
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
Qwen/Qwen3-8B revision b968826d9c46dd6066d109eabc6255188de91218; NF4 four-bit double quantization; LoRA rank 16, alpha 32, dropout 0.05 on q/k/v/o projections; AdamW learning rate 0.0001; batch 1, accumulation 4; seed 42; 64 updates over 256 selected arithmetic training questions (128 FinQA and 128 TAT-QA). The same data, actual order and optimizer budget are used for the matched text/fusion models.
Trainable parameters: 15,335,424. The text-only baseline has no numerical encoder or added prefix.
Every row below uses the same frozen questions and answer instructions. FinQA is raw execution-reference agreement at five decimals, not official program/execution accuracy or pure arithmetic accuracy. The fraction prompt conflicts with some reference percentage/index-point conventions; a separately frozen reference-only unit canonicalization diagnostic is retained in the full results, with ambiguous/inconsistent references flagged. TAT-QA uses its official EM/F1 scorer on arithmetic questions only. Each source has 64 test questions; these are not full-benchmark results. Synthetic labels cover 128 cases in 32 four-case scenarios; quartet accuracy requires all four labels to be correct.
| Model | FinQA raw reference (%) | TAT-QA EM (%) | Synthetic label (%) | All-four quartet (%) |
|---|---|---|---|---|
| B0 | 0.00 | 7.81 | 0.00 | 0.00 |
| B1_seed42 | 17.19 | 32.81 | 12.50 | 0.00 |
| N1_seed42 | 6.25 | 23.44 | 3.91 | 0.00 |
| P1_seed42 | 15.62 | 35.94 | 0.00 | 0.00 |
| B1_seed11 | 12.50 | 35.94 | 19.53 | 0.00 |
| N1_seed11 | 10.94 | 23.44 | 1.56 | 0.00 |
| P1_seed11 | 10.94 | 29.69 | 0.00 | 0.00 |
| B1_seed23 | 15.62 | 31.25 | 0.00 | 0.00 |
| N1_seed23 | 12.50 | 23.44 | 88.28 | 56.25 |
| P1_seed23 | 12.50 | 29.69 | 0.00 | 0.00 |
B0 = unadapted model; B1 = text QLoRA; N1 = numerical fusion; P1 = independently trained, number-independent four-token prefix control. P1 matches prefix positions, not N1's parameter count, and was added after development results but before test prediction. All three predetermined seeds are reported in EVALUATION_RESULTS.json, including paired uncertainty and secondary comparisons. No pooled cross-dataset score or state-of-the-art claim is made.
Across all three fixed seeds, restricted TAT-QA EM averages 33.33% for text-only and 23.44% for fusion; every seed favors text-only. Fusion's higher mean strict synthetic label score (31.25% versus 10.68%) is driven by one seed and marked format sensitivity. The separately labeled post-hoc scale-agnostic diagnosis instead averages 94.27% text-only and 89.06% fusion, with all cases retained. These outcomes do not establish a stable numerical-fusion reasoning advantage.
The initial seed-42 development comparison favored text-only adaptation: 6/64 versus 3/64 on FinQA and 23/64 versus 9/64 on TAT-QA. Zeroing numerical features changed no correctness outcomes; value permutation changed three with a net loss of one. These do not demonstrate a useful numerical-value contribution under that recipe. The paper distinguishes development, secondary diagnostics, replication and frozen test results. POST_HOC_SCHEMA_DIAGNOSTIC.json separately reports counterfactual label-only results while ignoring the scale string in otherwise complete single-label JSON. This was added after inspecting B0 test outputs, because its scale strings failed the strict schema; it keeps all cases in the denominator and does not replace the frozen strict scores.
Use the included loader so the backbone revision, four-bit settings, tokenizer, prompt schema and any numerical prefix match the tested recipe. The tested path requires CUDA.
python
from huggingface_hub import snapshot_downloadcheckpoint = snapshot_download("alinakhay/Meridians-8B-Text-QLoRA")print(checkpoint)From the downloaded directory:
bash
python -m pip install -r requirements.txtpython inference_meridians.py --model . --input example_input.jsonThe example is an input-format illustration, not a measured performance claim. Inputs contain source (finqa or tatqa, selecting the stated answer convention), narrative, a two-dimensional string table, and question. Complete prompts above 1,920 tokens or tables above 128 numerical cells fail explicitly. Thinking is disabled and decoding is greedy with at most 64 output tokens. The FinQA prompt requests fractional percentage values despite the documented mixed raw-reference convention; TAT-QA uses displayed values and an explicit scale.
The same flat-file Hub layout and the included standalone loader were used on the final evaluation worker for all seed-42 test predictions. CODE_SOURCES.zip contains the corresponding package source and license notices. EXPORT_MANIFEST.json records byte-preserved file hashes, and RELEASE_FILES.json covers the complete upload bundle. The paper, full metrics, frozen protocol and upstream data provenance are included. The backbone weights themselves are downloaded separately from Qwen.
One epoch on 256 selected arithmetic questions is a small adaptation study. Context selection, schema compliance, numerical precision, scale handling and the extra encoder's alignment all limit interpretation. Public-benchmark pretraining contamination is unknown. FinQA report groups used for selected train/dev are excluded from the restricted test; TAT-QA table UUIDs do not establish source-document disjointness. Synthetic quartets do not establish real-world generalization. The technical report is a draft, not peer-reviewed or submitted.
Backbone-derived adapters retain Apache-2.0 terms (LICENSE). Included original code is MIT (CODE_LICENSE_MIT.txt), with upstream scorer license notices retained in their packages. FinQA: Chen et al., EMNLP 2021; TAT-QA: Zhu et al., ACL 2021. Their dataset attribution/licenses remain separate from the original synthetic generator's CC0 declaration. Raw upstream corpora are not bundled in this model repository.
from huggingface_hub import snapshot_downloadcheckpoint = snapshot_download("alinakhay/Meridians-8B-Text-QLoRA")print(checkpoint)python -m pip install -r requirements.txtpython inference_meridians.py --model . --input example_input.json