Upload 7 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

adapter_config.json

@@ -0,0 +1,41 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen2.5-0.5B-Instruct",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.18.1",
+  "qalora_group_size": 16,
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "q_proj",
+    "v_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86fdf271def69f020af33f1172225c9d62855e617a11dd7607265c23282d347e
+size 2175168

final_metrics.md

@@ -0,0 +1,32 @@
+# Final Training Metrics — Heuristix Reflection LoRA (v1)
+
+## Base Model
+- **Model:** Qwen2.5-0.5B-Instruct  
+- **Quantization:** 4-bit (bitsandbytes NF4)
+
+## Dataset
+- **Reflection training samples:** 120 examples  
+- **Format:** Question → Initial Answer → Self-Critique → Revised Answer  
+
+## Training Configuration
+- **Epochs:** 3  
+- **LoRA Rank (r):** 8  
+- **LoRA Alpha:** 16  
+- **LoRA Dropout:** 0.05  
+
+## Resource Usage
+- **Peak VRAM usage:** ~2.8 GB  
+- **Total training time:** ~20.44 minutes  
+
+## Final Training Result
+- **Final training loss:** 2.278  
+
+## Notes
+This LoRA adapter was trained to induce **self-reflection behavior** in a compact
+language model. The training demonstrates that reflection-formatted supervision
+can be learned with **low VRAM usage and short training time**, making the
+approach feasible on consumer-grade GPUs.
+
+---
+**Status:** Day 5 complete ✔  
+**Project:** Heuristix Self-Reflective LoRA (Research Preview)

readme.md

@@ -0,0 +1,25 @@
+# HeuristixAI Self-Reflect Qwen 0.5B (v1)
+This repository contains LoRA adapters trained to induce
+self-reflective reasoning behavior in a compact language model.
+## Base Model
+Qwen/Qwen2.5-0.5B-Instruct
+## Method
+Parameter-efficient fine-tuning (LoRA) on reflection-formatted data:
+Prompt → Initial Answer → Self-Critique → Revised Answer
+## Capabilities
+- Structured reasoning
+- Self-critique behavior
+- Reduced hallucination
+- Improved logical consistency
+## Training Setup
+- LoRA r=8, alpha=16, dropout=0.05
+- 4-bit NF4 quantization
+- Dataset size: 120 reflection examples
+- Peak VRAM: ~2.8 GB
+- Training time: ~20 minutes (GTX 1650)
+## Usage
+See reflection_lora_v1_demo.py for example inference.
+## License
+Adapters released for research use.
+---
+Developed by HeuristixAI.

reflection_lora_v1_demo.py

@@ -0,0 +1,68 @@
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
+from peft import PeftModel
+
+BASE_MODEL = "Qwen/Qwen2.5-0.5B-Instruct"
+LORA_PATH = "heuristix_reflection_lora_v1"
+
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_compute_dtype=torch.float16,
+    bnb_4bit_use_double_quant=True,
+    bnb_4bit_quant_type="nf4",
+)
+
+print("Loading tokenizer...")
+tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
+
+print("Loading base model...")
+base_model = AutoModelForCausalLM.from_pretrained(
+    BASE_MODEL,
+    quantization_config=bnb_config,
+    device_map="auto",
+    trust_remote_code=True,
+)
+
+print("Loading reflection LoRA adapter...")
+model = PeftModel.from_pretrained(base_model, LORA_PATH)
+
+print("Model ready!\n")
+
+
+def generate(prompt):
+    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
+
+    outputs = model.generate(
+        **inputs,
+        max_new_tokens=350,      # ↑ allow full reflection
+        temperature=0.7,         # smoother text
+        top_p=0.9,               # better continuation
+        do_sample=True,          # prevents early cutoff
+        repetition_penalty=1.1
+    )
+
+    return tokenizer.decode(outputs[0], skip_special_tokens=True)
+
+# ---- Interactive reflection demo ----
+while True:
+    user_q = input("\nEnter a question (or type 'exit'): ")
+
+    if user_q.lower() == "exit":
+        break
+
+    reflection_prompt = f"""
+Answer the question, then critique your answer, then give a revised answer.
+
+Question: {user_q}
+
+Format:
+Initial Answer:
+Self-Critique:
+Revised Answer:
+"""
+
+    result = generate(reflection_prompt)
+
+    print("\n=== REFLECTION OUTPUT ===")
+    print(result)
+    print("=" * 40)

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d429fe753aea0ff87a94e86396d5508abb0d1d0e1f7a0d47c787ff72e0bf2691
+size 11422170

tokenizer_config.json

@@ -0,0 +1,29 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "is_local": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}