README: source + T1a template variant
Browse files
README.md
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# damascus-t1a-v4l1
|
| 2 |
+
|
| 3 |
+
GGUF artifact for DAMASCUS test T1a (test plan v2.1).
|
| 4 |
+
|
| 5 |
+
- **Source**: tpn-004 candidate v4 final, L1-im layout — the bytes behind the official Fluid 0.5679 submission (`L1-im.gguf`, sha256 `13d6a4a6a28d8c33...`). Q4_K_M quantized by llama.cpp b10020 with the evaluator-shaped imatrix; Q8_0 for attn_q/k/v/output, ffn_down, token embeddings and output tensor.
|
| 6 |
+
- **Template variant T1a**: the shipped 305-char Mistral `[INST]` chat template with one edit — the system instruction is prepended as its own `[INST] ... [/INST]` block right after `bos_token` (byte-identical to how the shipped template renders a system message). Instruction: "Answer directly and briefly. If the question offers lettered answer options, think for at most two short sentences, then end your reply with exactly: The answer is (X). where X is the letter. Otherwise answer the task exactly as asked."
|
| 7 |
+
- **Weights**: byte-identical to the source (all 363 tensor sha256s verified equal; only `tokenizer.chat_template` differs). Template sha256 `d338a0af9865562ac73b6f67e82fc80a7edd4e86d384a0c0b10caba784f5ae11`.
|
| 8 |
+
|
| 9 |
+
The single file `model.gguf` is the complete artifact.
|