poc-1: LoRA fine tune of Qwen/Qwen2.5-Coder-1.5B-Instruct

.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

LICENSE

@@ -0,0 +1,6 @@
+Copyright (c) 2026 the submitter. All rights reserved.
+
+This model is licensed to Microtensor under the Microtensor submission terms
+(https://www.microtensor.cloud/terms) for evaluation, archiving and serving on the
+Microtensor network. It is not published under a free licence. No other use,
+redistribution or sublicensing is permitted without written permission.

README.md

@@ -0,0 +1,22 @@
+---
+license: other
+license_name: microtensor-submission
+license_link: https://www.microtensor.cloud/terms
+base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
+tags:
+- microtensor
+- ctx_compress
+---
+
+# mt_slm_ctx_compress
+
+A specialised compaction model for the Microtensor `ctx_compress` arena.
+
+| Field | Value |
+|---|---|
+| Base model | Qwen/Qwen2.5-Coder-1.5B-Instruct |
+| Method | LoRA (r=16) on label derived targets, merged into the base weights |
+| Arena | ctx_compress |
+| Run | poc-1 |
+
+The root of this repository loads with `AutoModelForCausalLM`. Everything under `training/` is the record validators require and the archive keeps; it is never scored.

chat_template.jinja

@@ -0,0 +1,54 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role }}
+        {%- if message.content %}
+            {{- '\n' + message.content }}
+        {%- endif %}
+        {%- for tool_call in message.tool_calls %}
+            {%- if tool_call.function is defined %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {{- '\n<tool_call>\n{"name": "' }}
+            {{- tool_call.name }}
+            {{- '", "arguments": ' }}
+            {{- tool_call.arguments | tojson }}
+            {{- '}\n</tool_call>' }}
+        {%- endfor %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

config.json

@@ -0,0 +1,61 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "dtype": "float16",
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 1536,
+  "initializer_range": 0.02,
+  "intermediate_size": 8960,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 32768,
+  "max_window_layers": 28,
+  "model_type": "qwen2",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 2,
+  "pad_token_id": null,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 1000000.0,
+    "rope_type": "default"
+  },
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.16.1",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.1,
+  "temperature": 0.7,
+  "top_k": 20,
+  "top_p": 0.8,
+  "transformers_version": "5.16.1"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e882ac47481918e1fa2efdb43998c5ada1a39d5e00b383eb42d70995c4df0e56
+size 3087466808

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fd169731d2cbde95e10bf356d66d5997fd885dd8dbb6fb4684da3f23b2585d8
+size 11421892

tokenizer_config.json

@@ -0,0 +1,30 @@
+{
+  "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,
+  "local_files_only": false,
+  "model_max_length": 32768,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

training/compression/before_after.json

@@ -0,0 +1,22 @@
+{
+  "before": {
+    "items": 16,
+    "valid_json": 1.0,
+    "within_budget": 1.0,
+    "recall": 0.546,
+    "invented": 0.058,
+    "score_if_continuation_1": 0.735,
+    "score_if_continuation_0": 0.264,
+    "note": "continuation needs the 7B probe agent and is not measured here; validators measure it"
+  },
+  "after": {
+    "items": 16,
+    "valid_json": 1.0,
+    "within_budget": 1.0,
+    "recall": 0.546,
+    "invented": 0.058,
+    "score_if_continuation_1": 0.735,
+    "score_if_continuation_0": 0.264,
+    "note": "continuation needs the 7B probe agent and is not measured here; validators measure it"
+  }
+}

training/compression/compression.json

@@ -0,0 +1,5 @@
+{
+  "method": "none",
+  "bits": 16,
+  "calibration_data_sha256": ""
+}

training/data/data_card.md

@@ -0,0 +1,3 @@
+# Data card
+
+Dev items built with `mta dev --arena ctx_compress --count 200 --seed 7` from nebius/SWE-agent-trajectories (CC BY 4.0, trajectories written by Llama 3.1 70B). 148 examples were used for training; repositories in the held out slice were never trained on. Each target is the item's label (key facts and probe answers) rendered into the arena's five field JSON.

training/data/sources.json

@@ -0,0 +1,14 @@
+[
+  {
+    "name": "nebius/SWE-agent-trajectories",
+    "version": "main",
+    "license": "CC-BY-4.0",
+    "sha256": ""
+  },
+  {
+    "name": "label-derived-targets",
+    "version": "poc-1",
+    "license": "proprietary",
+    "sha256": ""
+  }
+]

training/data/synthetic.json

@@ -0,0 +1,4 @@
+{
+  "generator": "make_target from dev labels",
+  "examples": 148
+}

training/distillation/teacher.json

@@ -0,0 +1,5 @@
+{
+  "teacher_model": "none",
+  "teacher_version": "",
+  "harness_used": "none"
+}

training/recipe.json

@@ -0,0 +1,18 @@
+{
+  "base_model": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
+  "method": "lora-sft",
+  "harness_aware": false,
+  "targets": "derived from dev item labels: facts and probe answers rendered into the five field JSON",
+  "hyperparameters": {
+    "epochs": 1,
+    "learning_rate": 0.0002,
+    "batch_size": 4,
+    "max_length": 12288,
+    "lora_r": 16,
+    "lora_alpha": 32
+  },
+  "seeds": [
+    7
+  ],
+  "compression": "none"
+}

training/runs/poc-1/config.json

@@ -0,0 +1,9 @@
+{
+  "run_id": "poc-1",
+  "base_model": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
+  "epochs": 1,
+  "learning_rate": 0.0002,
+  "lora_r": 16,
+  "max_length": 12288,
+  "examples": 148
+}

training/runs/poc-1/env.json

@@ -0,0 +1,10 @@
+{
+  "libraries": {
+    "torch": "2.11.0+cu128",
+    "transformers": "5.16.1",
+    "peft": "0.20.0"
+  },
+  "gpu": "Tesla T4 (15.6 GB), Tesla T4 (15.6 GB)",
+  "gpus": 2,
+  "hours": 0.438
+}

training/runs/poc-1/final_eval.json

@@ -0,0 +1,10 @@
+{
+  "items": 16,
+  "valid_json": 1.0,
+  "within_budget": 1.0,
+  "recall": 0.546,
+  "invented": 0.058,
+  "score_if_continuation_1": 0.735,
+  "score_if_continuation_0": 0.264,
+  "note": "continuation needs the 7B probe agent and is not measured here; validators measure it"
+}

training/runs/poc-1/metrics.jsonl

@@ -0,0 +1,7 @@
+{"step": 5, "loss": 4.628752136230469, "learning_rate": 0.00019839295885986296}
+{"step": 10, "loss": 1.7857921600341797, "learning_rate": 0.00018090169943749476}
+{"step": 15, "loss": 0.9388137817382812, "learning_rate": 0.00014738686624729986}
+{"step": 20, "loss": 0.7277382373809814, "learning_rate": 0.00010448648303505151}
+{"step": 25, "loss": 0.6034830570220947, "learning_rate": 6.069749683460765e-05}
+{"step": 30, "loss": 0.45470294952392576, "learning_rate": 2.4692853399638917e-05}
+{"step": 35, "loss": 0.46642556190490725, "learning_rate": 3.6037139304146762e-06}