Upload README.md with huggingface_hub

README.md

@@ -18,18 +18,22 @@ datasets:
   - mechramc/codek-v1
 ---
 
-# CodeK LoRA -- Qwen2.5-Coder-7B-Instruct
+# CodeK LoRA v0 -- Qwen2.5-Coder-7B-Instruct
 
-A LoRA adapter fine-tuned on the **CodeK v1** dataset: a reasoning-first, pedagogical coding dataset in Karpathy's register.
+A LoRA adapter fine-tuned on the **CodeK v1** dataset: a reasoning-first, pedagogical
+coding dataset. Teaches decomposition, bug diagnosis, contrast reasoning, and
+hypothesis-driven thinking about code.
 
-## What it does
+## v0 Eval Results (Pass 2 ground-truth, 50 seeds)
 
-Teaches the model to reason like an experienced engineer -- not just write code, but:
-- Decompose problems before implementing
-- Explain *why* a solution is the right one
-- Identify and diagnose subtle bugs
-- Contrast clean vs clunky implementations with precision
-- Apply a hypothesis->experiment->evaluate loop to code
+| Model | Pass@1 |
+|-------|--------|
+| Base (Qwen2.5-Coder-7B-Instruct) | 64% |
+| **LoRA checkpoint-800** | **58%** |
+
+6% regression on bug diagnosis. LoRA wins on 2/50 seeds (more direct, correct),
+base wins on 5/50 (LoRA misidentifies function or pattern-matches to training data).
+See `BASELINE_V0.md` in the dataset repo for full analysis.
 
 ## Training
 
@@ -37,25 +41,20 @@ Teaches the model to reason like an experienced engineer -- not just write code,
 |---------|-------|
 | Base model | `Qwen/Qwen2.5-Coder-7B-Instruct` |
 | Method | LoRA (RS-LoRA) |
-| Rank | 16, Alpha 32 |
+| Rank / Alpha | 16 / 32 |
 | Dropout | 0.05 |
 | Epochs | 3 |
-| Batch size | 2 (effective 8 with grad accum) |
+| Batch (effective) | 8 |
 | Learning rate | 2e-4 |
 | Train pairs | 2,351 |
-| Val pairs | 262 |
-| Final eval loss | 0.0600 |
-| Hardware | RunPod A100 80GB |
-| Training time | 59 minutes |
+| Best eval loss | 0.0583 (step 528) |
+| Checkpoint used | checkpoint-800 (eval loss 0.061) |
+| Hardware | RunPod A100 80GB, 59 min |
 
 ## Dataset
 
-[mechramc/codek-v1](https://huggingface.co/datasets/mechramc/codek-v1) -- 201 seeds, 2,613 ShareGPT pairs across 4 augmentation passes:
-
-- **Pass 1** -- Reasoning decomposition (intuition, plan, why this solution, alternatives)
-- **Pass 2** -- Debugging (introduce + diagnose a subtle one-line bug)
-- **Pass 3** -- Contrast (clunky vs clean, with specific explanation)
-- **Pass 4** -- Research loop (hypothesis, minimal test, success metric, simplicity check, abandon condition)
+[mechramc/codek-v1](https://huggingface.co/datasets/mechramc/codek-v1) --
+201 seeds, 4 augmentation passes, 2,613 ShareGPT pairs.
 
 ## Usage
 
@@ -67,12 +66,3 @@ base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct")
 model = PeftModel.from_pretrained(base, "mechramc/codek-qwen2.5-coder-7b-lora")
 tokenizer = AutoTokenizer.from_pretrained("mechramc/codek-qwen2.5-coder-7b-lora")
 ```
-
-Or with Unsloth (2-5x faster inference):
-```python
-from unsloth import FastLanguageModel
-model, tokenizer = FastLanguageModel.from_pretrained(
-    "mechramc/codek-qwen2.5-coder-7b-lora",
-    max_seq_length=4096,
-)
-```