GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
sulaimanalshammari
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
sulaimanalshammari
Model Tree
humain-ai/ALLaM-7B-Instruct-preview
Input Modalities
Output Modalities
Supported Functionality
The latest local re-audit supersedes the earlier 600/600 headline.
Frozen V6 symbolic-entailment holdout:
a2d27bc8a249c0d32fdd33c7abc89323d50b1000aba9a18b111bc9334d4c6a63The verified R3/D3 result file SHA-256 is:
6a136fbc61109dbfbf756962dfb5afebc4ed3967c14413aa13f8531b263b8a99
The verified B0 result file SHA-256 is:
11841e08ea8eb237ded20cd50d3525080f072b6ef95f92e62b91957735dc2b48
The actual final-file SHA-256 of the B0-vs-R3 comparison JSON is:
60f2e1453f74953fb2c8771315c59d5479f2c6ff55dc43f839b1dc13bd418632
An older evaluation artifact reported 600/600 for R3/D3. That result is superseded and must not be used as the current audited score.
The repository also retains task-specific V6 receipts for other models evaluated under their recorded binary scoring protocols:
| System | Correct | Accuracy |
|---|---|---|
| ALLaM-7B base (B0) | 357/600 | 59.50% |
| R3/D3 | 367/600 | 61.17% |
| Qwen3-8B | 430/600 | 71.67% |
| Mistral-7B-Instruct-v0.3 | 412/600 | 68.67% |
These numbers are not a general ranking of the models. Qwen used a single-token next-token binary logit gate. ALLaM/R3 and Mistral used their recorded sequence/log-likelihood protocols. Tokenizer structure and scoring implementations differ, so comparisons are limited to this frozen task and the documented receipts.
logic_augmented_v3_train.jsonl - 960 rowslogic_augmented_v3_dev.jsonl - 240 rowsed67fe365a29ebc7c3b1e7dda771fb44490c4a5a7bbbca4a0fe8bafe434e15c962c0ab46fe0c827b4a0d551635f4b022c4b0cd8536d353c5183547842dc86c06192The training manifest records v6_used_for_training=false, v6_used_for_scoring=false, and v6_errors_inspected=false before the final V6 gate.
r=16, alpha=32, dropout=0.05q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj1120181e-4, cosine schedule, 4 warmup steps0.69630696317569040.168248259136453280Adapter SHA-256:
c6a90800e51c3a8f8568cda42edbebf7334f9d81fbafb9a11e10a72b436d0fe7
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_id = "humain-ai/ALLaM-7B-Instruct-preview"adapter_id = "sulaimanalshammari/R3-D3-ALLaM-7B-Logic-Adapter" tok = AutoTokenizer.from_pretrained(base_id)base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")model = PeftModel.from_pretrained(base, adapter_id)model.eval()training_manifest.json - training configuration and source hashesR3_D3_VERIFIED_STATUS_20260930.md - audited statusbenchmarks/V6-R3-D3-REAUDIT-RECEIPT.json - re-audit synchronization receiptbenchmarks/r3-d3-allam-v6-seqgate.jsonl - verified adapter row resultsbenchmarks/r3-d3-allam-v6-seqgate-summary.json - verified adapter summarybenchmarks/r3-d3-vs-b0-allam-v6-comparison.json - audited base/adapter movementbenchmarks/V6-QWEN3-8B-COMPARISON-RECEIPT.jsonbenchmarks/V6-FOUR-MODEL-COMPARISON-RECEIPT.jsonThis release demonstrates a small measured task-specific gain over the ALLaM-7B base on the frozen V6 symbolic-entailment benchmark: +10 net correct answers, with both improvements and regressions.
It does not establish general reasoning superiority, broader Arabic-language superiority, or superiority over Qwen, Mistral, or other language models.
NO CLAIM BEYOND THE HASH.
Base model license: Apache-2.0. This repository distributes the adapter and supporting evidence, not the 7B base-model weights.
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_id = "humain-ai/ALLaM-7B-Instruct-preview"adapter_id = "sulaimanalshammari/R3-D3-ALLaM-7B-Logic-Adapter" tok = AutoTokenizer.from_pretrained(base_id)base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")model = PeftModel.from_pretrained(base, adapter_id)model.eval()