Update README with proper YAML metadata and formatting

README.md

@@ -1,3 +1,18 @@
+---
+license: other
+tags:
+  - distillation
+  - sft
+  - qwen3.8
+  - kimi-k3
+  - fable-5
+  - nanbeige
+  - coding
+  - reasoning
+  - tool-use
+  - 4096-context
+---
+
 # Qwen3.8-Kimi-Fable Distillation
 
 A curated SFT dataset distilled from three teacher models, designed for fine-tuning compact models like Nanbeige4.2-3B.
@@ -6,8 +21,8 @@ A curated SFT dataset distilled from three teacher models, designed for fine-tun
 
 | Source | Rows | Description |
 |--------|------|-------------|
-| Qwen3.8-Max Preview (Alibaba) | ~10,000 | 49,796 teacher traces across math, code, reasoning, and instruction-following. Domain breakdown: math 33%, code 28%, reasoning 24%, instruction 15%. |
-| Claude Fable 5 (Anthropic) | ~2,500 | Verified coding and debugging trajectories generated via moonshiner. Focus: tool-calling, multi-step planning, build/test/fix loops. |
+| Qwen3.8-Max Preview (Alibaba) | ~10,000 | Teacher traces across math, code, reasoning, and instruction-following. Domain breakdown: math 33%, code 28%, reasoning 24%, instruction 15%. |
+| Claude Fable 5 (Anthropic) | ~2,500 | Verified coding and debugging trajectories via moonshiner. Focus: tool-calling, multi-step planning, build/test/fix loops. |
 | Kimi K3 (Moonshot AI) | ~1,200 | Verified coding, tool-use, and instruction-following traces. Focus: parallel tool calls, state tracking, verification-driven completion. |
 
 **Total: 13,700 training rows**
@@ -27,10 +42,10 @@ Each row is a JSON object with:
 ## Quality Filters
 
 Applied during curation:
-- Quality score ≥ 7.5 on Qwen3.8 traces (median 9.0, mean 8.6)
+- Quality score >= 7.5 on Qwen3.8 traces (median 9.0, mean 8.6)
 - No refusals, no empty responses
 - Near-deduplication via SHA-256 hashing
-- Context length ≤ 3800 raw tokens
+- Context length <= 3800 raw tokens
 - Minimum response length: 30 tokens (non-trace), 15 tokens (trace)
 
 ## License