voidful/Qwen3.5-9B-gemini-3.1-opus-4.6-reasoning

• Qwen/Qwen3.5-9B
• DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT
• voidful/Qwen3.5-9B-gemini-3.1-opus-4.6-reasoning

---

Model Summary

• Base model: Qwen/Qwen3.5-9B
• Model type: Causal language model
• Primary focus: Reasoning, multiple-choice QA, and general zero-shot evaluation
• Strengths: ARC, BoolQ, aggregate benchmark performance
• Trade-offs: Slightly weaker than some baselines on HellaSwag and OpenBookQA

---

Evaluation Setup

• 0-shot
• No few-shot examples
• Same benchmark suite
• Reported with standard error

Compared Models

1. Qwen/Qwen3.5-9B
2. DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT
3. voidful/Qwen3.5-9B-gemini-3.1-opus-4.6-reasoning

---

Main Results

Representative 7-task Average

Macro Average over All 12 Reported Metrics

---

Benchmark Results

---

Key Observations

Strengths

• The model achieves the best overall average score across the compared models.
• The model shows clear improvements on ARC-Challenge and ARC-Easy.
• The model strongly outperforms DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT on BoolQ.
• The gains are especially visible on reasoning-oriented benchmarks.

Trade-offs

• The model is not the top model on every benchmark.
• The model is slightly behind DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT on HellaSwag and OpenBookQA.
• The model is close to the baselines on PIQA and Winogrande.

---

Interpretation

• structured answer selection
• reasoning-oriented multiple-choice QA
• calibration on science and reading-style benchmarks

---

Limitations

• The evaluation here is limited to a small set of common zero-shot benchmarks.
• Some benchmark differences are small and may fall within the reported standard error.
• The model should not be described as universally better on every task.
• Additional evaluations on instruction following, long-context reasoning, coding, multilingual performance, and open-ended generation are still needed.

---

Conclusion

• the best overall aggregate benchmark score
• the strongest ARC performance
• strong BoolQ performance
• competitive general capability on other zero-shot commonsense tasks

---

Raw Results

Qwen/Qwen3.5-9B

DavidAU/Qwen3.5-9B-Claude-4.6-HighIQ-INSTRUCT

voidful/Qwen3.5-9B-gemini-3.1-opus-4.6-reasoning

See axolotl config

axolotl version: 0.16.0.dev0

yaml

1# Example config for RCCA-TR A+ (Reliability-Calibrated Conflict-Aware Trust-Region) fine-tuning 2# A+ variant: only 1 model in GPU memory (active model) 3# Prior = offline cache, EMA = drift buffer 4 5base_model: Qwen/Qwen3.5-9B 6 7plugins: 8 - axolotl.integrations.rcca_tr.RCCATRPlugin 9 - axolotl.integrations.liger.LigerPlugin 10 11liger_rms_norm: true 12liger_glu_activation: true 13 14# Enable RCCA-TR trainer 15rcca_tr_trainer: true 16 17# Conflict score hyperparameters 18rcca_tr_conflict_lambda1: 1.0 # weight for surprisal in conflict score 19rcca_tr_conflict_lambda2: 0.5 # weight for margin-based conflict 20rcca_tr_conflict_tau: 1.0 # temperature for conflict sigmoid 21 22# Reliability score hyperparameters 23rcca_tr_reliability_beta: 0.5 # balance between stability and evidence 24rcca_tr_reliability_tau: 1.0 # temperature for reliability sigmoid 25 26# Trust-region hyperparameters 27rcca_tr_epsilon_min: 0.01 # minimum trust-region radius 28rcca_tr_epsilon_max: 1.0 # maximum trust-region radius 29rcca_tr_kl_lambda: 1.0 # Lagrange multiplier for KL penalty 30rcca_tr_use_smooth_objective: true # smooth g(r_t)*KL vs hinge 31 32# Drift buffer (replaces EMA model) 33rcca_tr_ema_decay: 0.999 # decay rate for drift buffer 34rcca_tr_drift_gamma: 1.0 # drift → reliability scaling 35 36# Prior cache (optional; omit to use fallback mode) 37# rcca_tr_prior_cache_path: ./prior_cache/prior_cache.pt 38 39# Dataset 40datasets: 41 - path: voidful/gemini-3.1-opus-4.6-reasoning-merged 42 type: chat_template 43 split: train 44 45dataset_prepared_path: ./prepared_data/rcca_tr 46 47chat_template: qwen3_5 48 49# Training settings 50sequence_len: 16384 51sample_packing: true 52pad_to_sequence_len: true 53 54gradient_accumulation_steps: 4 55micro_batch_size: 1 56num_epochs: 3 57optimizer: adamw_torch_fused 58lr_scheduler: cosine 59learning_rate: 2e-5 60 61bf16: true 62gradient_checkpointing: true 63flash_attention: true 64 65dataloader_num_workers: 0 66 67deepspeed: deepspeed_configs/zero2.json 68 69val_set_size: 0.05 70 71save_strategy: epoch 72 73output_dir: ./outputs/rcca-tr-fft 74 75hub_model_id: voidful/Qwen3.5-9B-gemini-3.1-opus-4.6-reasoning 76push_to_hub: true 77hub_strategy: end 78 79log_on_each_node: false 80logging_steps: 1 81

Qwen3.5-9B-gemini-3.1-opus-4.6-reasoning

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• learning_rate: 2e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 80
• gradient_accumulation_steps: 4
• total_train_batch_size: 320
• total_eval_batch_size: 80
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• training_steps: 6

Training results

Framework versions

• Transformers 5.3.0
• Pytorch 2.10.0+cu128
• Datasets 4.5.0
• Tokenizers 0.22.2